Skip to content

What is new in versions 3.6 to 4.0

3.8 improves two ways to work without the keyboard: the agent understands the screen more reliably, and voice conversation correctly recognizes the credentials you already configured.

4.5.8: search in the model and provider pickers

Section titled “4.5.8: search in the model and provider pickers”
  • Search for a model instead of scrolling. The composer’s model dropdown now has a search box at the top: type part of the name and the list filters instantly, ignoring case and accents. Enter picks the first result.
  • Same for providers and profiles. In the list of available engines and AI you can search by provider name, profile name or the model each one uses.

4.5.7: tidy watch alerts and no fake bubbles on mobile

Section titled “4.5.7: tidy watch alerts and no fake bubbles on mobile”
  • Watch alerts without flooding the chat. Matches from a watch are grouped: the first one arrives right away and the next ones are bundled into one alert every 2, 4, 8 minutes, up to 15. They come in as system notices, not as if you had typed them.
  • On mobile, internal notices no longer look like your messages. Watches, background processes, heartbeats and reports coming from the desktop are shown as notices instead of a user bubble (mobile app 2.0.9).

4.5.6: the agent drives TestApi end to end and picks a model per task

Section titled “4.5.6: the agent drives TestApi end to end and picks a model per task”
  • The agent builds and runs your API tests. Ask in the chat for a collection and it creates it in TestApi with its requests and tests, runs it and reports the result; everything it creates is saved and visible in the TestApi tab.
  • WebSockets in TestApi. WebSocket requests are saved in collections with their messages, open in their own editor, run alone or with the rest of the collection, and survive Postman import and export. The agent can also test them and bring back the conversation with the server.
  • Postman-style environments. New Environments section with global variables and the active environment; each collection has its variables, auth and scripts from its menu.
  • Model by level (experimental). With the composer’s «Model by level» switch on, each subagent gets the cheapest model among your providers that reaches its task’s difficulty and has enough context and quota. Each subagent report states the chosen model. Off by default.
  • Widgets without flicker. A widget loads once when created and only reloads if its content changes. /widget appears in the slash menu.
  • Remote connection. You can see which chats are working, long chats load fully, and subagents, pins and order show in the sidebar.
  • A project menu at the end of the list opens upwards instead of being cut off, and the context budget notice explains that history is compacted automatically.

4.5.5: turns that survive drops and a cleaner work thread

Section titled “4.5.5: turns that survive drops and a cleaner work thread”
  • A connection drop no longer kills the turn. If the network goes away mid-response, the agent resumes from where it stopped instead of treating it as finished or getting stuck on “I can’t answer”. Network drops and provider timeouts are retried automatically, and a tool call that arrived truncated is never run half-formed.
  • Long answers continue on every provider. Automatic continuation when the model runs out of output budget now works with Claude and Gemini too, not only OpenAI-compatible ones. Recent Anthropic models no longer reject the continuation.
  • End of turn without tallies. The final answer no longer carries the “Cierre: … OK: n” line or the automatic list of failed commands: if a failure matters, the agent says so in the answer itself. Incident tracking keeps running behind the scenes.
  • Cleaner work thread. Tool runs are described in past tense once they finish, an error the agent retried and fixed shows in neutral (“n retries”) instead of red, while working only the last two steps are shown with “n more steps”, and the “Working” and “Long turn finished” notices are gone. The task panel adapts to narrow widths.
  • More accurate computer use. Screenshots are scaled by model tokens, so on widescreen displays the coordinates the model sees match the real screen and clicks do not drift. Shortcuts that take the agent out of the approved app (Win+R, Alt+Tab, Win+D…) ask for approval, and a key chord cannot exceed eight keys.
  • Browser. Console and network buffers are cleared when the origin changes and the filter accepts regular expressions (/pattern/); tabs the agent opened close without confirmation while yours still ask; and with several browsers online the agent asks which one to use.

4.5.4: widgets, machine file index and faster local voice

Section titled “4.5.4: widgets, machine file index and faster local voice”
  • Floating widgets. Save a canvas as a widget and it floats over the app. Turn them on, off, rename or delete them from the Widgets menu.
  • Machine file index. In Settings › Privacy you can enable an encrypted index of all your drives that never leaves the machine. It refreshes once a day in the background and models use it to find files instantly. It skips caches, node_modules, AppData and system folders.
  • Faster real-time local voice. With Qwen3 voices, the reply starts playing when the first clause is ready instead of the whole sentence.
  • Your voice with Qwen3-TTS. Voices from “My voices” can also be used locally with Qwen3-TTS.
  • Qwen3-TTS now checks at startup that it runs on the GPU: machines that had it installed CPU-only repair themselves.
  • The end-of-turn summary no longer shows traces from exploration commands.

4.5.3: TestApi with the response side by side

Section titled “4.5.3: TestApi with the response side by side”
  • TestApi looks the way it should. The response sits to the right of the request, with a draggable splitter and a button to put it below if you prefer. The URL is readable with {{variables}} highlighted, methods and HTTP status use the theme colour (4xx/5xx errors and DELETE use the theme’s red), the response JSON has highlighting and line numbers, and tabs with unsaved changes show a dot. Breadcrumb “collection / folder / request” with “Save” next to it.
  • Lighter Windows installer: it no longer drags along development environments or leftovers from other platforms.

From 4.4.0 to 4.5.2: TestApi, Fish Audio, the mobile mailbox and an app that no longer freezes

Section titled “From 4.4.0 to 4.5.2: TestApi, Fish Audio, the mobile mailbox and an app that no longer freezes”
  • TestApi: your API client inside Context Code. New tab in the top bar (Context · Developer · TestApi · Database · Studio · Design · Forge). Workspaces, collections, folders and requests with tabs; Params, Headers, Body, Auth and pre/post scripts with the pm.* API and asserts; response with highlighting, headers, cookies and console; global, environment and collection variables; Postman v2.1 import/export, OpenAPI 3 import, history and code generation in 14 languages. The agent uses it as a tool too: paste an API’s documentation or a collection into the composer and it can build the collection, run the requests and export it to your project. Requests go through the same network approval policy as everything else. If you had GetMan, your workspaces are imported.
  • Fish Audio as an optional voice. Pay-as-you-go provider with your own key (encrypted, never sent to the frontend). Pick from its Spanish voices with search and preview, and use it both in “Read text” and in real-time voice, streaming, with automatic fallback if it fails.
  • The desktop no longer freezes in long chats. Two memory leaks in the message list: long autonomous turns mounted all their blocks with no cap, and the window expanded when scrolling up never shrank back. If it ever locks up, on reopening the owner account sees where the report is to send it.
  • Encrypted mobile mailbox. Leave tasks from your phone while the desktop is off: they travel end-to-end encrypted, the server cannot read them, and the desktop runs them on start-up and sends the answer back. Active only with a paired phone, with a setting to turn it off.
  • Computer use. UWP apps such as Calculator, window capture with Windows.Graphics.Capture (Electron windows no longer come out black), and lab tests no longer steal your focus.
  • Voice. Field test matrix per platform and automated smoke tests for dictation, reading and start-up without configuration.
  • iOS ready for TestFlight with a single command from the Mac, a plugin checker and native UI tests; browser extension 1.2.0 (inspection, pointer, zoom, file upload, GIF recording and PDF) on the Chrome Web Store.
  • Turn close-out without repeated lists or redundant counts; the CLI degrades gracefully on consoles without Unicode and ships a first-run guide.

4.3.0: agents that resume, computer use with per-app permissions, and alerts on your phone

Section titled “4.3.0: agents that resume, computer use with per-app permissions, and alerts on your phone”
  • Subagents you can resume and fork. An agent that is still running or working in the background can be continued with its context instead of relaunched; fork creates a child with the parent’s whole conversation, and each child can have its own effort level. Read-only delegations start without asking; write delegations ask once and work in their own worktree.
  • Less context spent on tools. The agent looks up a specific tool (BuscarHerramienta) instead of activating whole groups; MCP servers start on first use and disabled ones no longer appear in the index; the catalog always keeps the same order so the cache holds. Discovering a tool does not authorize it: permissions stay the same. The agent also sees how much context it has used and can drop groups it no longer needs, and gets a short correction when it repeats a costly pattern.
  • Computer use with per-app permissions. The agent asks for access to each application and you approve it; the browser is view-only and the terminal left-click only. Each chat owns its desktop while it works, screenshots hide passwords, the vault, notifications and unapproved windows, it never performs financial actions or opens links by clicking, and each step reports whether it ran, whether the screen changed and whether the expected result happened. A global shortcut stops everything, including in the background.
  • AskUser with several questions. Up to four questions in one card, with checkboxes, “Other” with text and a preview to compare options; with asynchronous questions the agent keeps working on what does not depend on your answer.
  • Memory that checks itself. Before using a memory, the agent checks that what it cites still exists and otherwise marks it as possibly outdated; the automatic extractor merges new findings into the similar entry instead of discarding them, and repeated failures of a tool are recorded as an incident in .context/incidencias/.
  • Accurate context panel. The header, the composer indicator and the sum of the breakdown rows show the same number, and the estimate includes the system prompt and the tools sent.
  • More rigor in every task. Verify decides what runs before running it and returns a typed result; the end of a turn carries a faithful report of what was done; compactions keep prohibitions and decisions; and external content (channels, large pasted text, pages, process output) comes in marked as data, not instructions.
  • In the background. PowerShell in the background, the Vigilar and Despertar tools, and “Stop” in the background work panel.
  • Browser extension 1.1.0. Per-site permission prompts and per-host browser permissions, dialogs that no longer hang the tab, screenshots that do not steal focus, the tab’s console and network, shadow DOM and same-origin iframes, and jobs with acknowledgement, expiry and cancellation.
  • Mobile. “Jobs” panel with Stop for background processes, subagents and loops; approve or deny from the notification; alerts via self-hosted ntfy/UnifiedPush; sync with an explicit cursor; and end of turn with a recap and buttons.
  • Comment on a selection in the chat, like in Telegram: the quote travels to the composer with its source.

4.1.0: loops, the agent reads any window, and every model reasons the way it should

Section titled “4.1.0: loops, the agent reads any window, and every model reasons the way it should”
  • Loops: work that triggers itself. In the project’s .context/BUCLES.md you declare a trigger (every new commit or every heartbeat), the steps, the verification and when to notify you; the heartbeat runs it. Every loop starts in “record only” mode (it logs in .context/bucles/ what it would have done, touching nothing) until you switch it to active. They show up in Settings › Heartbeat with their log.
  • Computer use: windows that used to be blind. Electron apps (ChatGPT, Slack, VS Code…) exposed four controls; the agent now wakes up their accessibility before reading (ChatGPT goes from 4 to 221 named controls). WinUI apps (WhatsApp, Settings) are read in 200 ms instead of timing out. And two new actions, snapshot and act, read the accessibility tree with stable references and act on them without coordinates or screenshots (with the optional agent-desktop library).
  • Each provider reasons with its own controls. Claude Fable, Opus 5 and Sonnet 5 failed with a 400 when reasoning was on; Kimi K3.1 ignored the level and always thought at maximum; DeepSeek could not be turned off; GLM-5.3, Grok, Qwen and MiniMax ignored the selector. The level you pick now reaches each one the way its API expects, and the selector warns when changing it mid-chat resets the cache. GPT-6 by API key uses the Responses API (the only one with tools), Kimi keeps its cache for an hour and does not lose it when tool groups are activated, and with Claude you can change effort without losing the cache and trim context on the server (option in Settings › Advanced).
  • Studio: every card says what it is waiting for. Awaiting your answer, awaiting review, blocked, queued, working or closed, derived from the chat’s real state; the ones that need a person are highlighted. A card you start yourself always has a slot, even when the plan keeps every worker busy.
  • Voice with Gemini: safety net. If the voice model answers “I can’t” without delegating, the request goes to the agent anyway and the voice tells you the result.
  • Resumed subagents check files, git and the outside world before repeating an action: their transcript may not reflect what actually happened.
  • Channels: a tunnel link’s lifetime can be 30, 60, 180 or 360 days (plus the short options), so you don’t have to rescan the QR every month; the mobile app warns you 5 days before it expires.
  • Mobile 2.0.6: when the link expires, «Reconnect» with a new QR (or by pasting the link) keeps the connection’s name, with no need to delete and recreate it; the app updates itself from Google Play in the background; a Copy/Quote/Listen bar over part of a message selected with your finger; and read-aloud can now be paused, resumed and skipped sentence by sentence.
  • Voice no longer says “I can’t”. In the real-time conversation with Gemini Live, OpenAI Realtime or Grok Voice, asking “open Excel” or “use the computer” could end in “I have no access”: the voice model believed its only tool was for code. It now speaks as a single agent in the first person (“on it”, “opening it”), knows exactly what your chat can do (computer use, browser, files, commands, email, databases, whatever you have connected) and, when in doubt, hands it to its engine instead of refusing.
  • It tells you what it is doing. While the engine works, the voice narrates the steps in its own tone (“looking at the screen”, “clicking Save”, “running a command”), without talking over you and without repeating itself.
  • Approvals by voice. If the engine needs your permission (in “Ask for approval” mode), the voice asks and a “yes” or “no” is enough; the desktop dialog closes on its own.
  • Channels: “Get stable URL” for the owner. With the owner session signed in, the button asked to sign in; the stable URL is now obtained with that same session.

4.0.8: Full access without interruptions, also in the background

Section titled “4.0.8: Full access without interruptions, also in the background”
  • Chats in “Full access” no longer wait for permission when they are not open. Before, if you were working on another project, the background chat received the request and stopped until you entered it (hence the notice in the chat list). Now it is approved instantly and silently, with the usual limits: never an elevated command, a phone call, a database destruction or a high risk.
  • Subagents of a plan follow the chat’s current mode. If you switched to “Full access” with the chat already open, the specialists kept asking with the old mode; now they use the live one.
  • Agent badge on mobile and on the web. The same desktop card (face, role, provider and model, what it is doing now, Assignment and Output) comes to the mobile app, one column on a phone and two on a tablet, and to the tunnel web, where tapping a badge opens the agent’s thread with “Back to the main thread”.
  • On mobile, agent adjustments are no longer painted as errors. When it calls a tool wrongly and corrects itself, it shows in grey with “Not your fault”; red is reserved for real failures.

4.0.7: agents introduce themselves with a badge

Section titled “4.0.7: agents introduce themselves with a badge”
  • Agent badge in the chat. Each subagent shows up as a card with its face, role, the provider and model it runs on, what it is doing right now in plain words (“Reading App.tsx”, “Running the tests”), how long it has been at it and how many operations it made. When done it shows “Work finished” and its report under Output; the brief it received is under Assignment. Clicking the card opens all its work. Several agents at once go in a three-column grid; the one you expand takes the whole row.
  • The main thread stays clean. What each subagent does (files it reads, searches, notices) no longer mixes with the main agent’s work: it lives inside its card and its tab. The context meter in the bottom bar no longer moves with subagent telemetry, which has its own window.
  • Errors that are not errors. When the agent calls a tool wrongly and corrects itself (a missing field, a format), it no longer shows “N with error” in red: it appears as an “adjustment” in grey with the note “Not your fault: it noticed and carried on”. Red is reserved for real failures (network, permissions, missing file).
  • Resuming a failed subagent opens a menu with the connected providers and the model (editable); “Resume” picks up where it left off. The resumption shows as a system line, not as a message of yours with code in it.
  • Permissions you can actually answer. If the app was closed with an approval open, the dialog no longer stays stuck when you come back: the stale request is discarded and the chat is free. Approvals asked by a subagent say which agent is asking.

4.0.6: approvals that do show up on the desktop

Section titled “4.0.6: approvals that do show up on the desktop”
  • Pending approvals show up on the desktop again. In long chats, after restarting or updating the app, a new permission request could stay hidden: the agent kept waiting and it could only be approved from the phone. Each request now has a unique id and the desktop always shows it. If two are pending, answering one shows the other.
  • The Vault does not ask in “Full access”: using, saving, rotating or testing a secret goes through without confirmation (the value still never goes through the chat). Only writing secrets to a .env asks, because it leaves the value in plain text on disk.

4.0.5: Jarvis-style voice and reverse engineering with Ghidra

Section titled “4.0.5: Jarvis-style voice and reverse engineering with Ghidra”

Real-time voice

  • Grok Voice (xAI’s grok-voice-3.0) as a third real-time conversation engine, next to OpenAI Realtime and Gemini Live. It needs an xAI API key.
  • Voice personality: Neutral, Jarvis, Friendly, Warm or Professional. It changes the tone, not what it can do.
  • Background tasks: what you ask by voice that needs real work is done by the chat agent while the conversation goes on; you can start several and the voice tells you each result when it is done.
  • ElevenLabs in “Your chat model” mode: it listens with Scribe and answers with the ElevenLabs voice, streamed, with the chosen personality too.
  • On the phone: Settings → Voice conversation lets you talk with the computer’s OpenAI Realtime, Gemini Live or Grok Voice, with personality and tasks, also away from home through the tunnel.

Reverse engineering

  • If the computer has Ghidra 12.1+ and a JDK 21+, the app registers the rea server (REA on Ghidra) by itself. Just ask “analyze this exe”: the agent opens the binary, decompiles it, follows strings and calls, renames what it understands and explains what it does. See Reverse engineering.

4.0.4: agents that learn from their own work

Section titled “4.0.4: agents that learn from their own work”
  • Improve an agent: in Agent Team, the sparkles button on each agent opens a chat that reviews its real tasks and proposes changes split into how it works and what it knows about its craft. It shows the proposal as a diff and applies nothing without your approval. Built-in agents improve through lessons; your team’s agents by editing their definition (the previous version stays in their history).
  • Lessons by kind of work: what the agent learns while doing a task is saved as a lesson with its “When” and handed to the subagent of the same kind before its next task, most confirmed first.
  • Lighter context: if the fixed instructions of each turn grow too large, the skills index and the graph summary shrink automatically (the agent can still request them). /contexto shows how much each part weighs.
  • Workflows with long outputs: a step’s full result is saved to a file and the next step receives an excerpt with the path instead of the whole text.
  • Safer skill and agent edits: all problems are reported at once; after two failed attempts it stops and asks you, and replacing a skill or an agent keeps the previous version.

4.0.3: history tied to your account and Listen to audio

Section titled “4.0.3: history tied to your account and Listen to audio”

History security

  • Your chats are tied to your account: the history is encrypted with a key that depends on your license account. If someone copies the file to another computer, it cannot be opened. On your computers it opens when you sign in with your account and a valid license; with an expired license it stays locked until the license is valid again, without losing anything.
  • Backup for moving to a new computer: in Settings → Security → History backup, Save backup… creates an encrypted file that only opens with your same account. Save it before formatting or switching computers; on the new one use Import backup… and restart. The previous history is kept as a fallback until the imported one opens correctly.
  • SQL console passwords encrypted one by one inside the history.
  • Key backup tied to the computer on Linux and macOS too, and the unencrypted copy left when migrating the history is deleted as soon as encryption finishes.
  • Notices and errors in this section appear in your language.

Listen to audio

  • New Listen to audio view (in the chats menu of the sidebar): upload a file or record from the microphone and it is transcribed with the local speech-to-text engine, with a timeline.
  • With the profile and model you choose it can improve the text or extract a requirement from what was said, live and with a Markdown view; copy it with one click or listen to it with the default voice.

Maintenance

  • Repair and Reinstall next to every installable component (voice, dictation, Edge voice, Context Forge, plugins, Laya and language servers), to fix a broken install without hunting for it.
  • The tool counter shows the tools the agent actually has when the router is on.

4.0.2: fewer tokens, Edge voice and model comparison

Section titled “4.0.2: fewer tokens, Edge voice and model comparison”

Faster and cheaper

  • Fewer tokens per request: the project context is no longer sent repeatedly, the MCP server list and the graph summary are shorter, and rarely used tools (social media, channels, vault, desktop, charts, automations) load only when the agent needs them.
  • Lightweight subagents: each subagent starts with the basic tools plus the ones its task requires, instead of the full catalog.
  • Prompt cache lasts longer: activating a tool group no longer invalidates it, the part that changes every turn goes at the end, and when a chat is paused the cache renews itself when the savings justify it (turn it off in Settings → Advanced).

Chat

  • /goal: set a goal for the session; the agent keeps it in mind every turn, even after compacting or restarting.
  • Per-chat approval mode: from the shield in the composer you choose how this chat approves actions without changing the global mode. When approving you can allow a pattern (for example pnpm test*) for the session only.
  • Compare across models: the same prompt on 2 to 4 models with answers side by side, and Continue with this one to keep going with the one you prefer.
  • /fork now really branches the conversation from the message you choose.
  • Background work in view: the top bar shows the subagents and processes still running, with an option to cancel them.
  • Ratings with a private note on every answer; never sent to the model.
  • Reasoning in your language: translate the model’s reasoning with one click or always, from Settings.

Tools

  • Issues and PR board now with Azure DevOps besides GitHub and GitLab.
  • CSV preview as a table and a safer Word preview.
  • Settings → Advanced gathers settings that used to require environment variables, with a «?» help on each field, and each project can have additional folders allowed for the agent.
  • Browser extension: dictation and read aloud, an «Ask Context Code» bubble when selecting text, attachments, multiple tabs as context and provider icons. It never reads password fields.
  • Real-time voice: interrupting also stops the agent’s work, and the Gemini conversation resumes on its own if the service drops it.
  • Read aloud: the Edge voice works again. Microsoft changed how its service is authenticated and the app now uses the new method.
  • More consistent agent: when it runs a long background process (a build, for example), it continues with the remaining steps of the task once it finishes instead of stopping to report.

Voices

  • Settings → Voice groups Voices, Reading, Dictation and Conversation into tabs.
  • Your own voices: record or upload a sample of your voice and use it with Chatterbox to read aloud.
  • Custom server: Reading accepts any OpenAI-compatible speech server (Kokoro-FastAPI, openedai-speech, AllTalk, LocalAI, Speaches), with URL, model, voice and optional key.

Account

  • Subscribe with PayPal from the plans window: you complete the payment in the browser and the license updates itself in the app once it is confirmed.

4.0.1: the changes card is back to its usual look

Section titled “4.0.1: the changes card is back to its usual look”
  • Turn changes: the files the agent changed are shown again in the usual card, with its icon, the expandable list and the Review and Request review buttons in view. When the turn had edits, a single card appears instead of two.

4.0.0: parallel work, automatic memory and providers without keys

Section titled “4.0.0: parallel work, automatic memory and providers without keys”

Agents and workflow

  • Workflow: the agent chains subagents with dependencies; steps without dependencies run in parallel, each output feeds the next one and the journal stays in .context/workflows, so an interrupted flow resumes without repeating finished steps.
  • Subagents in their own worktree: with worktree isolation each subagent works on a separate git branch; if nothing changed it is removed, and if something changed the report gives the branch and how to merge it.
  • More stable subagents: the swarm budget no longer counts cache reads, each subagent has its own token cap and, at 80 % of the time limit, it is told to deliver what it has instead of being cut off.
  • Agent settings in Settings → Sessions: token and time caps per orchestration, concurrent subagents, iterations per turn and more.
  • Read another conversation: the agent can consult another chat of yours, summarized (for a specific question or to continue the work), with credentials masked.
  • Deliverables and turn changes: when a task ends, one card lists the delivered files and another the changed ones, with View changes and Request review.
  • /fusion: the same question to four different models with the chat model judging the anonymous answers. /btw answers a side question without touching the conversation.

Chat

  • Incognito chats from the palette (Ctrl+Shift+P): no memory, no trace in search, deleted after 24 h.
  • Search: Ctrl+F inside the chat and # in Ctrl+P to search all conversations.
  • Comment selection from a message or the editor, sent as an attachment of the next message.
  • Rating with thumbs up or down on each answer and Quote another conversation from the sidebar.
  • /fork, Ctrl+Shift+T to reopen the last tab, chats waiting for your answer move to the top, Archive all, translate reasoning and hide whitespace changes in the diff.
  • Deny with a reason: when rejecting an approval you can write why, and the agent receives it.
  • The prompt queue keeps attachments when editing, the context meter shows the cache hit rate and the retry notice counts down the seconds.
  • Trajectory viewer (palette → “View the model trajectory of this chat”): every model call with its reasoning, tools and tokens.

Memory

  • Save what matters from my idle chats (off by default): a light model reviews idle chats and saves lasting facts; what it saved is listed in EXTRACCIONES.md. Incognito chats are never included.
  • Before compacting a long conversation, the facts that must not be lost are rescued into memory.

Providers and Context Forge

  • Vertex, Azure AI Foundry and Bedrock without pasting keys: the gcloud session, az login or the AWS profile are used.
  • Provider router: learns each profile’s real limits, tracks requests and tokens per minute and per day, probes the health of the automatic chain and orders the fallback by reliability, speed and quota.
  • Per-conversation prompt cache on Kimi, OpenAI, Codex, Qwen and DeepSeek, and Anthropic’s 1 h cache in the main chat.
  • Context Forge: official Gemma 4 template, faster prompt reading on MoE models, reasoning budget in Configure model and better VRAM estimates.

Tools and security

  • Browser extension: the agent works in your real browser with your sessions, from a sidebar with the Context Code chat; the tabs it touches go into a “Context Code” group. It can type a password you give it in chat or one from the vault, always with your confirmation and without seeing the value.
  • Issues and pull requests board in the Git panel (GitHub with gh, GitLab with glab, Azure DevOps with az) with Work on this, which opens a chat with the item.
  • Enterprise policy: a system file can block MCP servers and skills by allow or deny list.
  • VS Code themes imported from Open VSX with their editor and terminal palette.
  • Automations: a trigger script decides whether the task calls the model, retries on network or quota failures, and a NO_REPLY answer when there is nothing to report.
  • Per-project environment variables in Settings → Security, passed to the agent’s commands without the value reaching the model.
  • Read shows the schema of a SQLite database and warns about merge conflicts; WebFetch no longer reaches private networks; Approve for me detects destructive commands before estimating risk.
  • Live skills: skills installed mid-session are announced without restarting the chat.

3.9.3: ContextCode AI makes better use of tools

Section titled “3.9.3: ContextCode AI makes better use of tools”
  • Tools with ContextCode AI: when the model asks for a tool by its short name (for example list_allowed_directories from an MCP server) or for one that was not active yet in the turn, the app recognizes it and runs it instead of leaving the request written as text in the answer.
  • Order within a single step: if the agent writes a file and reads or searches it again in the same step, the read now sees the change.
  • A single ContextCode AI profile, shown as Context Code in the composer selector and without a Connect button in the provider list. If your session has no access, the notice explains that you need to sign in again or activate it with your account password.
  • 7-day free trial for new accounts.

3.9.2: ContextCode IA ready on sign-in and an assistant that takes you there

Section titled “3.9.2: ContextCode IA ready on sign-in and an assistant that takes you there”
  • ContextCode IA already set up: when you sign in, the provider is ready without pressing Connect or typing anything, and it is listed first in the composer selector and in Settings → Providers and models. If you were already using another provider, yours stays active. The session renews itself before it expires, so the models no longer vanish after a few hours.
  • Models from the catalog: the model list comes from the ContextCode IA catalog for your plan.
  • The assistant takes you to the screen: under each answer of the assistant (the ? icon in the top bar) you get Open in Context Code with the Settings section or the view being discussed: providers, voice, language, appearance, graph, SQL console, Context Forge, Studio, Design, vault or automations. One click and you are there.
  • “Automatic (Typesafe)” no longer shows in the composer selector until you set it up.

3.9.1: ContextCode IA turns on without a key

Section titled “3.9.1: ContextCode IA turns on without a key”
  • ContextCode IA from Settings: pressing Connect in Settings → Providers and models → ContextCode IA no longer mixes it up with Context Forge or asks for an API key. It shows “Included in your license · no API key”, the Enable ContextCode AI button sets the provider up with the default model and, if you have not signed in yet, tells you to do it from Account. Before, it showed “Provider contextcode-ia not supported in this flow”.

3.9.0: AI included, an orb your way and a complete orchestrator

Section titled “3.9.0: AI included, an orb your way and a complete orchestrator”
  • ContextCode IA: your license brings an AI provider ready to use, with no keys or separate accounts. It is the default provider when you have not set up another, with a daily token quota per plan, a Switch provider notice when it runs out and the /consumo command to see usage. See Providers.
  • Voice conversation with OpenAI Realtime and Gemini Live: both engines use every agent tool (files, commands, computer use) like any other model; Gemini now says the tool result instead of going silent, the connection retries, the turn is cancelled when you interrupt and half-spoken text is trimmed.
  • The orb, your way: three new characters (Aurora, Echo and Constellation), configurable width and height for the floating widget and your own wake phrases, up to five and in any language (“Oye Jarvis”, “Hey Context”…). The name in “What should I call you” is Context by default and is the one used in “Hey …”. See The voice orb.
  • Composer ear: turn the always-on listening on and off with one click, like the other buttons.
  • An orchestrator that manages its teams: besides creating teams it now lists, fixes and deletes them (one agent or the whole team) when you ask. When planning it takes into account the project graph, diagrams, MCP servers and installed skills, and hands those resources out per step. A crashed, stopped or out-of-tokens subagent resumes from the chat with another profile or model via Resume with…. See Orchestration.
  • ElevenLabs in Design: in the Music tab, sound effects, music, voice design, voice changer and voice isolation with your account. See Design.
  • ElevenLabs telephony: import your Twilio or SIP number, assign it an agent, use it in the chat and make a test call with your plan’s minutes. See Voice.
  • Faithful GitHub themes: palettes generated from Primer (the same source as the Visual Studio Code themes), with matching selection, scrollbars and terminal, plus high-contrast, colorblind and tritanopia variants.
  • Assistant inside the app: the panel floats at the bottom right, like on the web, and errors say why (no network, server down, timed out).
  • Mobile app: each project’s graph and diagrams can be viewed from the phone (APK 1.0.41).

3.8.8: in-app assistant, full chat history and tidier voice settings

Section titled “3.8.8: in-app assistant, full chat history and tidier voice settings”
  • Context Code assistant inside the app: the ? icon next to the gear in the top bar opens a small chat that answers from the Context Code documentation and links the pages it relies on. It is the same assistant as in the manual; it does not use your provider or your credits. Esc closes it.
  • The whole chat history: very long chats only showed the most recent messages. The chat still opens just as fast, and the rest now arrives in the background back to the beginning; if you scroll up earlier, it loads right away without losing your place. The turn dots on the right cover the whole chat and jump to any turn.
  • Tidier voice settings: three sections with clear names: Read aloud (engine and voices), Voice dictation (transcription engine and microphone) and Voice conversation (OpenAI Realtime, Gemini Live or your chat model). With your chat model, a summary shows which dictation and which voice it will use, with a link to each section.
  • With Gemini Live or OpenAI Realtime, your chat model does the thinking: the voice engine only listens and speaks; everything you say goes to the chat provider and model with its tools (files, commands, computer use) and the answer is read back. That lets you use Gemini’s voice with whichever model you want. The “Your chat model does the thinking” checkbox in Settings → Voice conversation lets the voice model answer on its own, as before.
  • Notifications no longer interrupt the conversation: while you talk to the app, system notifications are silent; the chime came in through the microphone and cut the agent off, with any engine.
  • ElevenLabs starts sooner: the conversation uses ElevenLabs’ fast model with your same voice, the first sentence starts playing at a comma without waiting for the model to finish it, and if you interrupt, the rest is not synthesized with another voice.
  • Snappier local dictation: the Nemotron or Parakeet model is loaded once and reused; every sentence of the conversation paid the full load (over a second) before transcribing.
  • More reliable voice conversation: interrupting the app while it speaks no longer makes it hear itself, long background noise does not open false turns, the chosen microphone is kept when switching engines, and errors (no key, no credits, no microphone permission) are explained clearly.
  • Up to 100,000 memories: the assistant can keep up to 100,000 memories per scope (1000 by default, adjustable in Settings → General). Each conversation gets the most recent ones and finds the rest by searching; saving stays fast with thousands of memories.
  • MiniMax M3.1 Flash (Preview): it shows up in the MiniMax model picker even if the public catalog does not list it yet, and “Refresh list” queries your account’s models again.
  • Invalid email at sign-in: if you type a username without ”@”, the app says so instead of showing “server error (HTTP 400)”.

3.8.7: voice conversation with your model and your voice

Section titled “3.8.7: voice conversation with your model and your voice”
  • Your chat model answers: in voice conversation, the same model you use in the chat replies, with its context and its tools.
  • Heard in the voice you choose: the reply is spoken with the voice set in «Read text». With an ElevenLabs voice it streams sentence by sentence, without waiting for the whole answer.
  • Listens with your dictation engine: the conversation uses the dictation engine you have configured (local Whisper or ElevenLabs Scribe), the same one as the text box microphone. If you never chose one, both use the same default.
  • ElevenLabs Agents is retired: you no longer need to create an agent in the ElevenLabs dashboard. If you had one configured, the conversation switches to your chat model on its own; your ElevenLabs key still works for voice and dictation.
  • Listening fixes: the conversation uses the microphone chosen in Settings (if it is no longer connected, it uses the default one) and no longer hears nothing when it starts before you have clicked the window.

3.8.5: free web search, ElevenLabs, extra devices and background Computer Use

Section titled “3.8.5: free web search, ElevenLabs, extra devices and background Computer Use”
  • Smarter, free web search: web search combines several general search engines with technical sources (Stack Overflow, GitHub, npm, crates.io and Wikipedia) depending on what you ask, puts official documentation first, avoids repeating the same site and can bring the relevant paragraphs from the top pages. It needs no key or subscription. If a search engine starts blocking, the app lets it rest for a while instead of insisting.
  • Editor .md files open in “Live”: opening a Markdown file in an editor tab shows it directly in Live mode, with Preview and Source one click away.
  • Give the assistant a name: in Settings, “What do you want to call me” lets you give it a name of its own (for example, “Jarvis”). The assistant introduces itself and answers to that name; leave it empty and it stays Context Code.
  • ElevenLabs as an optional paid voice: with your own ElevenLabs key you can use its voices for read-aloud, its Scribe engine for dictation and its agents for real-time conversation on the desktop. In the mobile app, read-aloud and dictation can use ElevenLabs through the desktop, without typing the key on the phone. It is a paid cloud service: it consumes credits from your account. More in Dictation, read-aloud and translation.
  • Extra devices: if your plan has reached its device limit, you can request one more for 10 USD per month or 100 USD per year (depending on your plan period) from the app notice or from My account. The request stays pending until the payment is confirmed; once approved, click “I already paid / revalidate” and the device is activated. See Extra devices.
  • Safer, more faithful Markdown viewer: any HTML inside a document is shown sanitized, the file’s line endings (Windows or Unix) are preserved and, if the file changed on disk while you had it open, the app does not overwrite those changes when saving.
  • Background Computer Use on Windows: in Settings → Computer use → Working mode you choose how the agent works. With “On the window” (recommended) it acts on the task window without moving your mouse or bringing it to the front, and you keep working; if an action has no effect that way, it asks once per session for permission to use the real mouse. “Real mouse and keyboard” is the usual behavior. “Hidden desktop” (experimental) opens the agent’s programs on an invisible desktop and the chat shows its progress with a live preview refreshed every few seconds; it only supports classic desktop programs (not Store apps), and switching modes or closing the app closes the programs it opened. On macOS and Linux the real mouse is always used.
  • The agent remembers the whole chat: when a long conversation is summarized to save space, the agent can look up the earlier messages again if it needs them, instead of losing them. And searching your past conversations now finds the most recent chats too.

3.8.4: live MD viewer and better-fitting diagrams

Section titled “3.8.4: live MD viewer and better-fitting diagrams”
  • Live MD viewer: Live mode, the one the viewer opens with, reads like a finished document and stays editable. Syntax only appears on the line where the cursor is. Tables render as tables, front-matter properties as a card, checkboxes toggle with a click, and code blocks are no longer mistaken for headings or bold text.
  • Theme colors: the MD viewer uses the app theme and reads well in light and dark. Preview has the same width and font as Live.
  • Reading position: when you return to a document, it opens on the same line where you left it.
  • Diagrams in Spanish: with the app in Spanish, the diagram viewer is in Spanish too (it was always in English).
  • Right height: the title and the whole diagram fit in view without shrinking too much; in the chat, the card grows up to 640 pixels to fit the diagram.
  • Present: opens the diagram full screen and Esc exits.
  • More readable nodes: node names and subtitles are larger when they fit in the box.
  • Diagrams use the available space: in the chat and after Expand, the diagram takes the full width. Before, it stayed in a narrow column (about 960 pixels in the chat) and left half of a wide tab empty.
  • Subtle scrollbar: inside the diagram the scrollbar is thin and matches the theme, instead of the thick, light system scrollbar.

3.8.2: one-click Diagnostics and MCP servers that repair themselves

Section titled “3.8.2: one-click Diagnostics and MCP servers that repair themselves”
  • Diagnostics in Settings: in Settings → Diagnostics, Check everything tests the AI provider, license, Node.js, npx, uv, Python, Git, every MCP server, skills, plugins, the data folder and free disk space. Anything that fails tells you what it found, why it fails and what to do, with a link to the Settings section that fixes it. Copy report gets it ready to send to support.
  • Python MCP servers missing a package: if a server dies with “No module named …” (for example an Arduino server that needs pyserial), the app relaunches it with its bundled uv and the missing package, without installing anything into your Python. If there is a requirements.txt next to the script, it uses all of it.
  • First install ready to work: built-in plugins are active from the first launch (with no account connected: you connect your own) and the packages the built-in MCP servers need are downloaded in the background after installing or updating, so the first conversation does not wait for them.
  • Up-to-date model picker: with your Anthropic or OpenAI account, the list comes from the provider’s own API: retired models disappear, new ones show up first and the rest keep their date order. A network error no longer leaves the picker on the fallback list for a whole day.
  • Remote MCP servers with sign-in: the token renews itself and open chats use the new session without opening another chat.
  • Python on Windows: servers launched with python start even when python is not on the PATH (the py launcher or uv’s Python).
  • Cleaned-up MCP catalog: servers whose job the app already does show as “Superseded”, in amber, instead of failing in red. The desktop-control one shows a short notice that fits the card.
  • “Node.js missing” with Node bundled: fixed a PATH that was too long and made cmd.exe miss freshly installed npx commands.
  • Edge TTS without Python: “Read aloud” no longer asks for Python or the edge-tts package, which becomes an optional fallback.
  • Diagrams: opening Diagrams from the side panel no longer switches to Developer mode; it only opens the tab.
  • Git: a GitHub Desktop-style remote bar with the right action (push, pull, sync or check origin).
  • SQL console: the console chat shows the connection permissions next to the input box.
  • Stable URL with your account: web access gets the machine’s fixed URL with your Context Code session, with no tunnel broker key.
  • Pinned projects no longer move on their own: before, the pinned project you were working on jumped to the top. Now they keep the order you choose.
  • Reorder them by dragging: drag a pinned project onto another pinned one and it takes that position. A newly pinned project goes to the end of the pinned list.

3.8.0: precise Computer Use and repaired voice conversation

Section titled “3.8.0: precise Computer Use and repaired voice conversation”
  • The agent understands the screen before acting: every screenshot includes the accessible controls in the active window, with their names and positions. It can click by name, type in a field, choose a menu, or fill in a form without guessing coordinates.
  • It verifies every action: it checks the area around the click, warns when nothing changed, and shows a close-up so it does not repeat blindly. If you move the mouse while it is working, it stops instead of fighting you for control.
  • Repeatable tasks without stored secrets: it can remember and replay action sequences. Passwords, keys, and card details are never written to the task and are requested again when it runs.
  • Sharper captures and cross-platform support: up to 1568 pixels per side, Windows UI Automation through COM, native Linux X11, and an experimental macOS mode that stays off until you enable it.
  • OpenAI Realtime finds the right key even when the last profile used is an OAuth account. The app searches every profile for a valid API key, distinguishes “account connected” from “API key available”, and refreshes the state when you connect a provider.
  • Gemini Live is selected automatically when OpenAI has no compatible key, and now uses the recommended gemini-3.8-live model. Local mode remains available without a provider.
  • Design only took up the left panel column (about 335 pixels) and the rest of the window stayed empty. The workbench now stretches across the full width when the projects drawer is closed.

3.7.1: new voice conversation, redesigned Design and a memory that learns

Section titled “3.7.1: new voice conversation, redesigned Design and a memory that learns”
  • Real-time conversation, for real: Conversation mode leaves LiveKit behind and talks directly to OpenAI Realtime or Gemini Live. Replies arrive in under a second, you can interrupt, and echo is handled by the browser itself. The voice model chats with you and, when you ask for real work (reading code, running something), hands it to your chat’s agent and reports back. Your API key never leaves the app. There is also a local mode with no voice provider that uses your dictation and your “Read aloud” voice. Pick it in Settings → Voice → Real-time conversation. You need an OpenAI or Gemini API key; signing in with the account does not work for voice.
  • Design, like a desktop generation app: Image, Video, Music, Gallery and Models tabs on top; tasks on the left (Text → Image, Edit, Upscale, Remove background, Image → Video…); a panel with the model, the prompt, the options and Generate always at the bottom; and the result shown large with a strip of everything generated in the chat. The Models tab shows the full catalog with filters and a details card to pick or configure. Projects open from the header.
  • A memory that learns, not just remembers: every entry accumulates proof with dated quotes; saving over an existing one reinforces it instead of overwriting, and if something contradicts it, it is marked doubtful until verified. Credentials are masked before saving. New knowledge pages: standing answers to fixed questions (“how do we ship a release?”) that the model writes from memory and git, and rewrites only when their sources change.
  • Vault: the secure field the agent asks for now actually shows up. Before, the agent said “I left you the field for five minutes” and nothing appeared. Now the box opens wherever you are and, if you close it, you find it in the Vault panel under Waiting for you to enter while it is still valid.
  • “Working in” for everyone: the Local / Server selector is no longer owner-only and is available with any active or trial license.
  • pnpm tauri:dev opens the window again instead of the terminal chat, and /voz also works from the tunnel web.
  • Remove links you no longer need: in Channels → Web access → Shared chats, revoked or expired links now have a Remove button that takes them off the list. Before, they stayed visible for up to a week. A link that is still active cannot be removed: revoke it first, so that removing never cuts off access by accident.
  • Allow works when working on another machine: connected to a machine over SSH or HTTP, clicking Allow, Always allow or Deny did nothing and showed a session_not_found error, because the answer was looked up in the local session instead of being sent to the machine where the agent runs. It now reaches the remote machine and the agent continues. The same happened with the agent’s questions (AskUser).
  • If the remote machine has “Allow approving permissions from the web” turned off in Channels, you will see that reason instead of a generic error.
  • The graph only reads your code: it now respects .gitignore when it is built. Before, it also walked ignored folders (build output, clones of other repositories, data) and, in large projects, they filled the symbol cap and left your own code out: the agent asked the graph, didn’t find the function and fell back to Grep. After updating, regenerate the project’s graph with /graph.
  • The agent searches the graph first: to locate code (where a function is defined, who calls it, what a change would break) it uses GraphQuery before Grep and Read, which cost many more tokens. Grep is left for literal text, such as messages or configuration keys.

contextcode can now manage a headless server without ever opening the window.

  • Remote access like in Channels: contextcode serve --public uses the machine’s fixed URL, and requests it on its own if you are signed in. Only if it can’t does it fall back to a temporary tunnel, warning that its URL changes on every start. --stable forces the fixed URL. With --print-qr you get the QR of the public URL and, below it, the local network one.
  • The QR no longer drifts from the service: serve --install prints, when it finishes, the URL and the QR with the same token the service will use. contextcode conexion --qr shows it again with the service already running. Starting a second serve on the same port stops with a clear message.
  • Paired phones: contextcode dispositivos lists them and dispositivos revocar removes their access. The revocation takes effect immediately even with the server running; before, the server could accept a revoked phone again.
  • Full management: mcp, skills, canal configurar/canal estado, automatizaciones, caja, modelos and uso, with the same data as the app.
  • Keys that don’t stay in the history: provider login --key, canal configurar and caja guardar ask for the value without showing it on screen, or read it from standard input.
  • Simpler Design, now with music: three tabs (Image, Video and Music), starter examples when the box is empty, and the technical settings (presets, seed, negative prompt, the model that enhances the prompt) folded into Advanced options. In Music you describe the style, choose the length (30 s to 3 min) and add lyrics or keep it instrumental; it is created in the cloud with ACE-Step through fal.ai. Guide: Design.
  • The agent sees your terminals: it reads what the project’s open terminals show without asking, and, always with your approval, types into them or stops them with Ctrl+C. With remote access, the panel terminal opens on the other machine. Details in Built-in terminal.
  • contextcode on Windows, no surprises: the help shows contextcode instead of the .exe name, which, when called directly, returns the prompt before it finishes and makes chat unusable. The built-in terminal and the agent’s shells also find contextcode even if the app was already open during installation. What to check if it doesn’t respond: If contextcode doesn’t respond on Windows.
  • Agents from private GitHub repositories: in Import agent you can pick a GitHub token saved in the project’s Vault. The value never goes through the interface or the AI provider and is only sent to GitHub. How: Importing agents from a private repository.

New guides: Remote access from a server and Manage from the terminal.

  • Copy and export diagrams on Windows: in a diagram’s Export menu, “Copy diagram” and the downloads (PNG, SVG, share card) did nothing on Windows. They now work, just like on macOS.
  • More reliable /diagrama: the command no longer gets lost silently when the chat is busy; it joins the running turn or tells you why it could not be sent.
  • Stopped plans: the stopped-plan banner now counts failed steps as resumable and no longer appears when there is nothing left to do.
  • Diagrams on your phone: from mobile app 1.0.38, the diagram card has a View diagram button that opens it in the light theme with pinch to zoom.

3.6.26: diagrams in the chat, local decisions and an agent that remembers

Section titled “3.6.26: diagrams in the chat, local decisions and an agent that remembers”

Ask “draw the architecture of this project” or “make a sequence diagram of the login” and the diagram shows up in the chat itself, interactive: zoom, search, light or dark theme and export. There are five types: architecture, workflow, sequence, data flow and lifecycle. The agent checks the project graph first so the architecture is the real one, and the engine makes sure arrows don’t cross or hide labels; if something is off, the agent fixes it on its own.

Diagrams are saved in the project and have their own Diagrams menu in the sidebar, next to the graph. You can also start one with /diagrama in the composer. The engine ships inside the app: nothing is downloaded and it works offline. Each diagram is shown isolated from the rest of the app and without network access.

Typesafe with Laya: decisions on your machine, for free

Section titled “Typesafe with Laya: decisions on your machine, for free”

The Typesafe AI card (Settings → Providers and models) now has a Who decides selector: Laya, which runs on your machine with no key and no cost per decision, or Jev, the usual cloud service. Laya installs with one click from the same card, with progress and a Cancel button (about 5 GB with an NVIDIA GPU, about 2 GB CPU-only), starts on its own with the first decision and stops when you close the app. If Laya is not ready, Jev decides when you have a key; otherwise the usual rules apply. In the chat, the decision line says when Laya decided. Jev is still the default.

An agent that remembers and works on its own

Section titled “An agent that remembers and works on its own”
  • Memory that searches inside your notes: before each turn the agent recalls the memory fragments about what you just wrote, without spending a model call on it.
  • Intents: “when we talk about the installer again, remind me to check who signs the package”. It fires when the topic comes up, not at a time.
  • Memory consolidation outside the turn, with no language model: it only reorders what is already written, it never invents.
  • Heartbeat with active hours, a per-project LATIDO.md checklist and its own model; it also works in contextcode serve without duplicate alerts.
  • Tasks: background work (processes, subagents, automations) is recorded and the agent can check or cancel it instead of relaunching it.
  • Live widgets (show_widget): the agent opens an isolated tab that updates while it works.
  • Meetings: the agent joins Meet, Teams or Zoom as a guest, transcribes who said what and leaves a summary in .context/reuniones/. In agent mode it can also speak.
  • Calls through Twilio: phone alerts and turn-based conversation, always with your approval and an allowlist of numbers.
  • The phone as a node: with the app open, the agent can ask it for a photo, its location or to show a notification.
  • Local design: upscale and remove background with ONNX models on your machine, with progressive preview, cancel and queued batches; real steps, guidance and scheduler in Forge Media.
  • Vault v1.1: guided rotation, generator in the wizard, importer for the manual vault, member removal with re-encryption, canonical repository lookup, trusted MCP switch and «secreto:id» expansion in .env with confirmation.
  • Live team chat: chats shared with the team sync instantly through a relay, with a “live” indicator and team management (create, invite, remove) from the sidebar.
  • Composer effort as a compact button with a slider.
  • The viewer follows the end of .log files, opens 7z and rar, and recognizes HEIC.
  • The Word editor keeps lists, headers, footers and merged tables.
  • The event trace is JSONL with secret redaction, retention and a disk cap, and warns when a child process keeps growing in memory.
  • LongCat 2.5 Preview as LongCat’s default model, with its published prices.
  • The sidebar is never narrower than the two-column menu.

The first time you open the app, a five-step guide gets it working: language, connecting the AI (with your account, with an API key or with a local model that downloads itself), choosing a folder and testing voice. Every step can be skipped, and you can see it again from your account menu.

There is also a permanent status icon in the top bar: it tells you whether anything is missing before the agent can answer — connected AI, valid model, internet, helper programs — and every warning comes with a Fix it button that takes you to the missing step. And chat errors now come out in plain language, with the action that solves them.

Real-time conversation — enable it in Settings → Voice and reading and press the conversation button in the chat. The agent greets you with “I’m listening.” and from then on:

  • It listens while you speak: nothing to press to start or to end the sentence.
  • You can cut it off mid-word and keep talking; it goes quiet instantly.
  • Live captions of what you say and what it answers.
  • If you go quiet, it asks “Are you still there?” instead of hanging up.
  • It answers with the voice you choose — Edge, Groq, Kokoro, Pocket or Chatterbox — and you can change it without hanging up.
  • When the answer includes code it doesn’t dictate it: it says “I’ll leave the code on screen” and writes it.

It works in the application’s 17 languages and also from your phone, over your home WiFi network.

And there is nothing to download: the Windows, Linux and macOS installer carries the voice agent, the server and the models inside, so Conversation mode works right after installing — even with no internet connection — without the first-time wait the preview builds had.

A floating orb, always on top of other windows, to follow the conversation while you work on something else. It shows the status (listening, thinking, speaking), a caption line and its own buttons to pause, return to the window and hang up. Drag it wherever you want and it remembers where you left it. You choose whether it appears never, when the main window is minimized — which includes closing it to the tray — or whenever voice is active.

You can give it three faces: Orb, Ring (three thin rings spinning) or Pixel (a 7×7 grid that fills up with your voice), with a live preview in the selector. And if you’re not sure it will show up, the Try widget button shows it for five seconds and writes the diagnosis underneath: whether the window was created, whether the system considers it visible, where it ended up and how many monitors it sees. If the saved position falls outside the connected screens, it returns to the center-right on its own.

“Hey Context”: have it listen without touching anything

Section titled ““Hey Context”: have it listen without touching anything”

With always-on listening you don’t even need to press the button: you say “Hey Context” — or “Oye Context”, or “Context” — and the conversation opens on its own, with the widget in front of you.

It ships off, and on purpose: it’s the only feature that opens the microphone without you asking. While it listens there’s an indicator next to the voice button that pulses with the microphone level, so there’s never any doubt. Everything happens on your machine: an energy detector discards silence without transcribing anything, and only the short bursts where someone speaks go through a small Whisper — about four tenths of a second of CPU each, and nothing at rest. With noise it demands hearing it twice in three seconds (three times if you say only “Context”), waits eight seconds before firing again, and never listens while a conversation is open or the agent is speaking.

Now the agent drives the app for you while you talk: “create a new chat”, “open project Invoices”, “switch to model Gemini”, “read me the last answer”, “pause”, “hang up”, “show the window”. The full list is on the widget page.

New xKiro provider: a gateway with more than 110 models from GPT, Claude, Gemini, DeepSeek, Qwen, GLM, Kimi, MiniMax or Grok behind a single key — and more than 40 free models on its free tier.

Gemini API Key no longer goes through the OpenAI compatibility layer: it uses Google’s native API. In practice:

  • “Model overloaded” 503 errors are retried on their own (2, 5 and 10 seconds) instead of cutting your turn short.
  • The Reasoning switch really controls Gemini 3’s thinking level.
  • Errors arrive in your language and with a hint of what to do.

Database and Developer: right where you left it

Section titled “Database and Developer: right where you left it”

Switching sections no longer costs you the work you’ve done: the SQL console and Developer mode keep the connection, the explorer tree, the open tabs and the cursor of every script. Plus:

  • Transparent reconnection and heartbeat for SQL Server and Oracle: if the server closes the session, the app reopens it without bothering you and retries once the query that failed because of the network.
  • Batches per engine: GO in SQL Server (with GO 5 to repeat) and / in Oracle, respecting strings and comments; errors tell you which batch and which line of the script.
  • Scripts that touch several databases on the same engine: SQL Server prepends the right USE and warns if the script names another database, with a button to rewrite it.

And you decide what the agent may do on each database

Section titled “And you decide what the agent may do on each database”

Every connection now has its Agent permissions button, with three presets — Read only (the factory default), Development and Administrator — and eleven individual switches: read, view schema, write, delete, create and alter objects, DROP, run procedures, transactions, export to file and ask before every write. Plus a cap on affected rows without asking.

It’s not a warning: what you switch off never reaches the database. The assistant checks it before every query and explains it in plain words (“I don’t have permission to delete on this connection; enable it in Agent permissions”); what does need confirmation tells you why and how many rows it’s going to touch, and the irreversible is always confirmed. A mixed script with a single forbidden statement is denied as a whole.

The phone stops being a viewer and becomes a command post:

  • Conversation mode with the same experience as the desktop: spoken greeting, live captions, waves that follow your voice, pause and resume, and Retry.
  • Voice commands that run on the phone: “create a chat called X” takes you to the new chat on the phone; “open the chat about X”, “switch to model Sonnet”, “read me the last answer”, “pause” and “hang up” are handled right there. The ones that only make sense in front of the PC are forwarded to the desktop, which confirms them back to you.
  • Attach photos and files from the paperclip, including the option to take a photo.
  • See the files the agent touches: in every read, write or edit operation the path appears; tap it and it opens in a read-only viewer.
  • The red square (“Stop the turn”) cuts the turn and its subagents, and clears whatever you had queued.

The APK is downloaded from the downloads page, with a QR code to install it from the phone itself. Conversation mode still needs to be on the same WiFi network as the PC (not through the tunnel) and the desktop turned on.

The terminal stops being the poor relation

Section titled “The terminal stops being the poor relation”

Context Code has always been one program with two faces, but from the terminal it was hard to know where you stood. 3.6 fixes that with three new commands — and a documentation section of its own:

  • contextcode conexion (also info or whoami) answers “where is everything?”: version and executable, whether you’re desktop+CLI or CLI only, the configuration and data folders, the database, the log, the remote link status and the paired phones. It shows no secrets: the URL with token only appears if you ask for it with --enlace or --qr. With --json, the same for a script.
  • contextcode instalar-cli puts the short command in the PATH without asking for administrator, right where the installer can’t do it: AppImage, macOS and the Windows .msi. (The .deb and the .exe installer already did.)
  • contextcode remoto and contextcode chat --remoto: chat from your terminal with the Context Code on another machine, the one at home from the office or the one on the server from your laptop. It pairs like one more device — exchanging the QR token for one of its own, which doesn’t expire — and, the important part, approvals are answered in your terminal: the agent over there doesn’t sit waiting for someone in front of that machine.

If you used an SSH workspace and it cut out every so often, the culprit was the home NAT or the VPN, which recycle idle sessions because they can’t see SSH’s encrypted traffic. Now there’s keep-alive on both layers, the app detects the dead tunnel and goes back to working locally on its own (instead of chaining three-second waits per command) and remote chats withstand outages of more than two minutes before giving up.

The new version downloads inside the app, with a progress bar and megabytes, and is verified with SHA-256 before running anything (“Checking the installer…”). You can also See what’s new without leaving the app, or leave it for Later. If something fails, Retry; the manual download link stays as the last resort, not the first option. On Linux .deb it gives you the command already written and copyable; on AppImage and macOS the handover is automatic. All the details, in Updates. (The intermediate step of choosing between “Install now” and “Install on exit” didn’t last long: 3.6.6 removed it — see below.)

Opening a chat with a long history took between 1.5 and 2.6 seconds; now it’s around 0.4. And coming back to the chat from Developer or from Database no longer freezes the interface.

The results grid behaves like a spreadsheet

Section titled “The results grid behaves like a spreadsheet”

Select rows with a click, Ctrl+click and Shift+click, all of them with Ctrl+A, and Ctrl+C copies as TSV ready to paste into Excel. The context menu adds copying with headers, as CSV or as INSERT. Context statements (USE, SET, DECLARE) no longer open a result tab, every “Resultado N” tab has its X and there’s “Close all results”. If the result came trimmed, it warns you before you copy half a report without noticing. See Database.

A chat that has delegated a lot piles up subagents in the sidebar. Now they can be deleted: one with the X, or from the parent chat “Delete finished subagents” and “Delete all subagents”. The ones still running aren’t deleted lightly — you’re offered Stop and delete or Just stop. See Agents and orchestration.

The widget now uses the application’s theme, the caption fits in two lines, and with “Show widget: when the main window is minimized” it appears when you minimize or hide the window even with no conversation running, with a “Talk to Context Code” button to start one. The voice detector is also stricter: fewer turns opened by noise. And no more echo: it no longer plays twice.

The “Context Code server” card disappears from Settings (it was a support thing). And the mobile app 1.0.31 drops the dictation button — the phone’s keyboard already does it better — and fixes the “missing chat parameter” error when opening the voice conversation.

Chats now open as tabs in the same bar as the files and the SQL console —“Project Invoices - Review the login”—, with a green dot when the chat is working and a ”!” when it is waiting for your answer. You jump between conversations with one click or with Ctrl+Tab, and activating a tab also takes you to that chat’s project (explorer, editor and Git). Closing it stops nothing: the chat keeps working in the background, and it tells you so the first time. Tabs are remembered when you reopen the app, and the “Chats as tabs” switch in Settings → General brings back the old behaviour. See Your first chat.

Pinning and archiving project-less chats too

Section titled “Pinning and archiving project-less chats too”

Right-clicking any chat in the sidebar —including those on the “Chat” tab, which had neither before— offers “Pin chat” and “Archive chat”. Pinned chats stay at the top and survive a restart; archived ones wait in their “Archived” section, and the search finds them with a “Restore and open” button.

Two labelled tabs with a counter, Changes and History. History brings the author’s avatar, relative time in your language, branch and tag labels, a search by message/author/hash, “Select branch to compare…” and “Load 100 more commits”; the commit detail, a copyable short hash, +N −M and the files with their status. And diffs are shown inside the panel with the Monaco editor: minimap, change markers in the scrollbar, line numbers on both sides, side-by-side or unified, folded unchanged regions and “Open in the editor” —with a simple fallback viewer for files over 2 MB—. In Changes, one click shows the diff and a double click opens the file in the editor. Along the way, a bug that made single-line commits disappear from history was fixed. See Git panel.

3.6.5: browsing without filling the bar, and Git with no surprises

Section titled “3.6.5: browsing without filling the bar, and Git with no surprises”

A single click on a chat in the sidebar no longer opens a new tab: it shows it in a provisional tab, with the title in italics, which the next single click reuses in the same place. That way you can look at ten conversations without ending up with ten tabs. The tab stays as soon as you do something that is not a peek: double click (on the chat or on the tab), typing a message or dragging it. You can turn it off in Settings → General with “Open chats as a preview with one click”. See Your first chat.

In the Git panel’s History, commits that are not on the remote yet carry a ↑ pill (“Not pushed yet”), and above the list there is a strip with “N commits to push · Push” and “M to pull · Pull”. If the branch has no remote it says so —“This branch has no remote · Publish branch”— and marks nothing, because there is nothing to compare against.

The diff in its own tab; the files as a tree

Section titled “The diff in its own tab; the files as a tree”

Diffs are no longer squeezed into the panel column: a click on a file —in Changes or in a commit— opens (or updates) the editor’s diff tab, now with a minimap, change markers in the scrollbar, folded unchanged regions and side-by-side ⇄ unified. The tab is reused; with Ctrl+click you open a separate one. A folded preview, off by default, remains in the panel. And the file lists —a commit’s and the Changes one— become a folder tree in an accordion, with a count and +/− per folder, single-folder chains compacted, a Tree/List switch that is remembered, collapse/expand all and keyboard control with ↑ ↓ ← → and Enter. See Git panel.

The optional update notice had been split into two steps —“Download” and, a while later, “Version X ready to install” with “Install now” and “Install on exit”—, which forced you back to the card for a click nobody had asked for.

It is the usual card again: “New version X available” with Update & restart, Later and the ✕, and See what’s new as a small text link. That single button does the whole way in the same card: download with progress (“120 / 389 MB”), SHA-256 check (“Checking the installer…”) and installation (“Installing and restarting…”), just like the mandatory ones.

Nothing underneath changes: the latest version is re-checked before downloading, on failure there are Retry and Open in browser, and whoever installed via .deb still stops at “Version X ready to install” with its copyable command —that package is installed by the system, not by us—. See Updates.

3.6.9: the graph understands your code, not just your files

Section titled “3.6.9: the graph understands your code, not just your files”

The code graph stops being a map of files and becomes a map of symbols: which function calls which, what inherits from what, and what crosses the border between the front end and the back end.

  • Questions that used to cost half a dozen searches. Now it is enough to say “explain handle_livekit_join to me and what breaks if I change it”: the agent answers with a single query that brings back the whole function, who calls it, what it calls and the files that depend on it. Behind it there are four new queries — explore, symbol, callers/callees and impact — with an impact radius up to depth 4.
  • Real TypeScript imports. They are read from the syntax tree, and the tsconfig.json aliases, the barrels (index.ts) and the re-exports are resolved: an import { X } from "@/core" stops counting as an external dependency and points at the real file. Inheritance is drawn too (extends, implements, impl Trait for, Python base classes), and C, C++ and C# are now parsed like the rest.
  • Front↔back bridges. Every invoke("…") in the front end is joined to its #[tauri::command] in Rust — and the ones pointing at a command that no longer exists (buttons that do nothing) are listed in the GRAPH_REPORT.md report. HTTP routes from Express, NestJS, FastAPI, Flask, ASP.NET and axum become nodes with their handler.
  • In the Graph view there are new types in the legend and the filters: Inheritance, Implements, Tauri invoke, Route → handler and the HTTP route nodes (shortcut 7).
  • And dependencies no longer eat it: when the graph’s caps are handed out, your code goes first and the vendored folders after, in an order that stays stable across regenerations.

Typesafe AI: the small decisions stop being guesswork

Section titled “Typesafe AI: the small decisions stop being guesswork”

Many of the app’s internal decisions —which model to use for the turn, which skills to load, whether an action is delicate, which specialists a plan needs— were made from hand-written keyword lists. With a Typesafe AI key they are actually decided, and with a confidence number: below the threshold the old rule comes back, so a decision is never worse than the one you already had.

  • The turn’s model when the selector is on Automatic, and each orchestrator subagent’s model: a mechanical task drops to the fastest, cheapest of your profiles, a redesign moves up to the most capable one. If the agent’s .md already declares profile: or model:, your file wins.
  • Relevant skills and risk in “Ask me first”.
  • The intent of what you write (do you want an explanation, or the work done?) and the domains it touches, which is where a plan’s proposed team comes from.
  • When the decision is its own, you see it: “Model chosen by Typesafe: … (confidence 0.87)”.

It is optional and cheap ($0.042 per million input tokens; a month of work is cents). With no key, or if the service fails, everything behaves exactly as before.

Go to definition now really works: Ctrl+click or F12 on a function, class or variable takes you to where it is defined, Alt+F12 peeks at it in a pop-up without leaving the file, and Shift+F12 lists the references. It is tested against real C#, TypeScript and Go servers; for C# two things had to be fixed that left the server silent (it never got an answer to its configuration request, and the solution path arrived with a Windows extended-length prefix). If the language has no server installed, the app says so and offers to install it. See Editor.

F5 no longer lists things that are not your project: it detects what is in the folder (.NET, Node, Python, Rust, Go, Java…), groups it by project and shows it with the language icon instead of the runtime name. One line per configuration, Ctrl+Shift+B for build targets, and everything can be pinned in .context/run.json. See Run.

Typesafe chooses with numbers, and tells you what it spends

Section titled “Typesafe chooses with numbers, and tells you what it spends”

Typesafe’s automatic mode no longer decides with adjectives: every candidate model carries its real price per million tokens and the measured speed of your previous turns, and profiles that just got a 429 or 503 drop out of the list while the quota lasts. The line under the message now reads “Model chosen by Typesafe: … · among 12 models, 2 rate-limited”. The Settings card gets a monthly limit in USD (once exceeded, every function switches off and you are warned once) and a “This month: N decisions · tokens · USD” line; the spend also shows in the Usage panel. The welcome guide has an optional step to paste and test the key, and without a key the “Automatic (Typesafe)” row shows disabled with a hint on where to set it up. From the phone and the tunnel web you can also put a chat on automatic and see the decision, and in the terminal contextcode chat --model auto works.

SVG with live preview, and a data console you can stop

Section titled “SVG with live preview, and a data console you can stop”

.svg files open in their own tab: the code on the left (with Format for XML that comes in a single line) and the live preview on the right, with a checkered background, zoom, drag, real size, Ctrl+S to save, an invalid-SVG warning and Export PNG at 1x, 2x or 4x. In Database, durations are readable now (“1 m 35 s 525 ms” instead of “95525 ms”), while a query runs the bar says “Running… 0:05” with the previous grid dimmed and per-statement progress, and there is a Stop button next to run (Alt+Pause or Escape) that really cancels on SQL Server, PostgreSQL, MySQL and SQLite and leaves the run marked “Cancelled by the user”.

Key Vault: your secrets never reach the AI

Section titled “Key Vault: your secrets never reach the AI”

New Key Vault panel (padlock next to Skills and MCP). Store passwords, tokens, SSH keys and server data (host, port, user and its password or key) encrypted with AES-256-GCM in .context/secretos/equipo.enc, with the master key in the system keychain. The agent asks permission to use a secret, never for its value: the key is injected into the command, the HTTP header or the trusted MCP server without going through the model, and everything that comes back from a tool is masked before it reaches the AI, so echo $TOKEN, env, cat .env or git remote -v with the PAT in the URL return «secreto:id» instead of the value. If you paste a credential into the chat, a banner offers to store it and send only the reference; when you open a .env or an appsettings.json with secrets, the editor tells you. With a strength validator (a key like Ds910913@ is flagged as weak), a generator, “Test connection” for servers, a usage audit log and an encrypted emergency export.

When the open project is Context Code’s own source code, the agent knows it: it cannot close or kill the instance it runs in (not even with Full access), and to test changes it uses pnpm tauri:demo, which starts a second isolated instance with its own configuration, the title “Context Code Demo” and an amber badge so you never confuse it with yours. Settings → General shows the Self-build card with versions, PID and the button to open the demo.

Subagents with the profile or model you say

Section titled “Subagents with the profile or model you say”

“Launch a translations agent with the DanielSalamanknf profile” or “with GLM” now works: the subagent tool and every step of the orchestrator plan accept profile (by id, readable name or provider) and model, with the list of your connected profiles visible to the agent. What you ask in the chat wins over the agent’s .md, and that wins over Typesafe’s automatic choice.

An installed version has 2 days from the moment the next one is published to update; after that it becomes mandatory. The notice tells you how many days are left. This 3.6.19 is the first one that applies it.

Share a chat: with a guest or with the whole team

Section titled “Share a chat: with a guest or with the whole team”

Two new ways to share a conversation, each with its own owner:

  • With someone outside, through the web link. In the chat menu, Share this chat… generates a link (and its QR) that opens only that chat: you choose View only or Can write, how long it lasts (1 hour, 1 day, 7 days or until you revoke it) and, optionally, the person’s name. Whoever enters sees the conversation and, if allowed, can keep working; their messages show as “Ana (guest)”. The agent’s approvals keep showing only on your desktop, never to the guest, and Key Vault secrets stay masked. Everything else in the app (other chats, projects, providers, files) answers “out of scope”. Live links can be seen and revoked from the chat or from the Connections panel.
  • With your team, through the repository. Share with the team saves the chat in .context/chats-compartidos/ as a one-line-per-turn file, with each turn’s author, and from then on every new turn is appended on its own. After a push, the rest of the team sees it in the project’s Shared by the team section, greyed out until they press Activate: then they follow it with their own provider and model, and what they write goes back to the repo. Git merges lines from two branches without conflicts (merge=union is set up automatically). No server, no internet, nothing exposed; any credentials found are redacted on export.
  • Much less memory with several chats open. Each chat started its own full set of MCP servers (files, git, SSH, email…) and kept it alive until the app closed: with six chats there were six copies of each one. Now each server starts only once and every chat in the project shares it; only the Playwright browser is still one per chat. A server shuts down when the last chat stops using it.
  • Stop now really stops what you launch with F5 in Developer mode. It used to close only the first process (for example cmd or pnpm) and left alive the program started underneath it (dotnet, node, the dev server), which kept holding the port. Now the whole process tree is closed.
  • If the administrator forces an update on your computer, the notice shows up again even if you had dismissed it.
  • No more stray processes after closing the app. Everything Context Code starts (MCP servers, voice, language servers, tools, local models) now closes with it, even if it crashes or restarts to update. Before, they could stay alive for days and one of them could exhaust the computer’s memory until Windows started showing “0xc0000142” errors when opening git, cmd or where. What you open outside the app (browser, Explorer, terminal, editor) stays open when you close it.
  • If a helper program fails to start, the tool gets the error right away instead of waiting behind a Windows dialog, and each command launches fewer intermediate processes.
  • Chat tabs are wider (up to 320 px) and shrink evenly when there are many: the project and chat name are now readable.
  • Settings → General no longer shows the “Self-build” card: it appeared in every project. The agent’s rules over its own code, the Demo badge and pnpm tauri:demo are unchanged.
  • From the tunnel web you can now attach images and files: a clip next to Send, camera on the phone, drag and drop onto the chat and paste with Ctrl+V; large photos are downsized before uploading and the limits (8 files, 12 MB) are warned about before you exceed them.
  • When sharing with the team, if .context/ is ignored in the repo the app tells you the exact rule and does not share halfway.
  • The agent heartbeat no longer enters your working chat: it uses its own “Heartbeat” chat, defers if you are typing and its “nothing to report” does not stay in the history.
  • Retry waits for Anthropic, OpenAI and Codex (429, 529 “Overloaded” and 5xx) now show on screen with the cause, the wait and the attempt, as they already did for Gemini.
  • In the Git changes list the folder is truncated and the file name is always visible, like in GitHub Desktop.
  • Archiving a chat from the Chats tab no longer sends you to Projects: you stay on the neighbouring chat of the same list, and the tab only changes when you open a chat of the other kind.
  • Duplicate in the file explorer’s context menu: it creates the copy and leaves it ready to rename.
  • With Gemini API Key, when Google throttles your quota (429 or 503) the retry shows on screen with the wait and the attempt number, and the error explains it is the key’s quota. HTTP connections are reused between requests.
  • MCP servers started with npx or uvx use the bundled runtimes, and arguments containing < or > no longer break the command on Windows.
  • The Conversation button can no longer crash the application.
  • Multi-cursor in the chat composer, like in the code editor.
  • The database object explorer scrolls horizontally.
  • The Linux Server card on the downloads page gets its installation instructions back.
  • The development build isolates itself from the installed app on its own.
  • Windows, macOS and desktop Linux: the application itself notifies you, downloads it and verifies it — see Updates.
  • Linux servers: contextcode update, or the same curl … | bash from the installation — see Context Code from the terminal.
  • Android: the app notifies you of the new version and takes you to install it.