codai for VS Code

AI that does the work and checks it. Follow your GitHub Copilot Chat sessions from your watch, phone and desktop — and answer them from there.

When Copilot's agent stops to ask "Allow run_in_terminal?" or shows you a question, you no longer have to be at your desk: the ask shows up on your Wear OS watch, the codai phone app and the codai desktop app. Tap Allow, Deny or pick an answer, and it lands in VS Code.

Features

• Shared sessions — every Copilot Chat session you work in (last 2 hours by default) is mirrored to your codai account with its title, prompts, tool calls and final answers.
• Approve tools remotely — pending tool confirmations appear as asks on the watch and phone. Allow / Deny are applied to the right chat in VS Code.
• Answer questions remotely — Copilot's question carousels appear with their options; pick one or dictate a free answer.
• Send a follow-up / stop — a prompt or a cancel sent to a session from another device is applied to that chat: the extension opens and focuses the target session first, so it never lands in whichever chat happened to be focused.
• @codai in chat — ask the codai model directly; /sessions lists your sessions on every device, /asks what is waiting for you.
• Pending prompts — the codai activity-bar view lists ready-to-paste prompts agents left for a future chat, grouped by project (the folder they were created in; prompts without one go under Other), then Due / Scheduled, with a badge of due ones. Open one in Copilot Chat (prefilled, not sent), copy, show, snooze, mark done, dismiss or delete it; save any selection with codai: Save selection as pending prompt. Handoff prompts (a ```text block in an answer that mentions a new chat / handoff) in mirrored sessions are saved automatically (codai.prompts.autoCapture).
• Prompt tracking, automatically — when you send a pending prompt in any Copilot Chat on this machine (every window and workspace, codai.prompts.scanAllWorkspaces), it is marked sent, then answered when the response finishes, with the chat, the turn and the git commit it started from. Matching is by text: exact (ignoring whitespace/case), the first 4000 characters, or ≥ 85 % word overlap; texts under 20 characters are ignored. No more "mark launched" by hand (still available as codai: Mark prompt launched (manual), or when codai.prompts.autoDetect is off).
• Verify — Verify prompt (inline on each prompt) and Verify all prompts (view title) ask codai whether the work was actually done. For prompts sent on this machine the extension sends the chat transcript of that turn and the git diff since it started (committed + uncommitted + untracked file names, each capped at 40 000 characters); others are verified server-side. The result opens in codai: Prompt report: per prompt the verdict (done / partial / not started), summary, what changed and what remains, plus totals and every remaining item in one list. The tree shows the state (sent / answered / verified) with icons, and the report summary in the tooltip.
• Transcripts — a second view in the codai activity bar lists your voice-recording transcripts (codai desktop, phone, watch) by day, with duration, source and processing status. Expand one to see its summary and the agent prompts codai suggests from it; Open transcript shows summary, key points, actions, open questions, suggested prompts and the full speaker-labelled transcript as a read-only markdown preview. Insert as prompt puts a suggested prompt or the summary into the Copilot chat input (not sent) or the active editor; Send to codai Prompts turns it into a pending prompt. Shown only when Transcripts is enabled for your account.
• Realtime — the prompts view follows the server live (server-sent events, reconnecting with backoff); a 5-minute refresh remains as a fallback.
• Web search and deep research for any model — three agent tools, #codaiWebSearch, #codaiWebFetch and #codaiDeepResearch, give Copilot agent mode verified web search, readable page fetch and cited multi-page research through your codai account — the same tools, with the same descriptions and results, as the codai desktop, phone and MCP server. They work with every model, including custom endpoints / BYOK models that have no built-in web search. Every search result was fetched and its snippet found on the page; published dates are shown when known. Search can be limited to or exclude domains (include_domains / exclude_domains), filtered by publish date (freshness, published_after, published_before), return longer verified excerpts (content_chars), and understands operators (site:, -site:, filetype:, intitle:, inurl:, "exact phrase", -word, after:2026-01, before:2025-06-30, OR). When codai's own results are weak, a separate Grounding with Google Search block is added (or ask for it with mode: google). Language and market are inferred from the query unless the model sets lang / country. #codaiDeepResearch (paid) returns a report where every claim cites its source; long jobs return a job id the model collects later. Failures come back as one clear error line (e.g. rate limit with the retry time), never as a silent fallback. Turn the tools on or off in the chat Tools picker.
• codai Studio for any model — eight agent tools create media and documents through your codai account: #codaiStudioMusic (song or instrumental, from $0.04), #codaiStudioImage (~$0.04), #codaiStudioVideo (short clip with sound, ~$0.10/s), #codaiStudioRender (HyperFrames HTML composition → MP4, $0.01/s, draft half), #codaiStudioSfx (free CC0 sound effects), #codaiStudioDocument (editable PPTX / DOCX / XLSX, free), #codaiStudioDocumentPreview (PDF + page images, $0.002/page) and #codaiStudioDocumentVideo (deck → narrated MP4). Paid tools ask before running and show the price. Results are saved in the open folder under .codai/studio/ (the first save offers to add it to .gitignore); with no folder open the tool returns a signed URL valid for 1 hour.
• codai models in Copilot Chat — codai: Add codai models to Copilot Chat adds the codai provider to VS Code's chatLanguageModels.json, so codai shows up in the Copilot Chat model picker.
• In your language — every message, picker, input box and tree label is in English or Romanian, following VS Code's display language.
• Experimental (Insiders) — sessions running on your other codai devices inside VS Code's own Sessions view (codai.experimental.chatSessionsProvider).

Get started

1. Install the extension and run codai: Sign in from the Command Palette.
2. Pick Continue with GitHub (uses the GitHub account already signed in to VS Code; codai: Sign in with GitHub does this directly), or Sign in in the browser and approve the code shown there (it is copied to the clipboard). No password is entered in VS Code; the resulting key is kept in VS Code's secret storage.
3. Open Copilot Chat as usual. The status bar shows codai N — the number of sessions shared — and a bell when something waits for you.
4. Optional: run codai: Add codai models to Copilot Chat (offered once after sign-in in VS Code and Insiders) to pick codai in the Copilot Chat model picker. It downloads the codai provider entry from https://ai.codai.ro/vscode/chatLanguageModels.json and adds or updates it in your user chatLanguageModels.json, keeping every other provider; the first change backs the file up as chatLanguageModels.json.bak-codai. Your key is never written to the file: VS Code asks for it the first time you pick a codai model and keeps it in the OS keychain.

Try without an account

Ask @codai anything before signing in and pick Chat free without an account (or run codai: Chat free — no account): 3 guest chats a day, 10 messages each, with the codai model. VS Code solves a short proof-of-work (about a second) instead of asking for an account. Guest chats are not stored by codai — the history stays in VS Code — and when you sign in later the last guest chat (up to 50 messages, at most 7 days old) is saved to your new account once. Shared sessions, /sessions, /asks, prompts and transcripts need a free account. By sending a guest message you accept the Terms and Privacy Policy and confirm you are 18+.

Compatibility

Requires VS Code 1.95 or newer. Some features depend on APIs only VS Code itself has; elsewhere the extension shows a one-time notice listing what is unavailable.

Details: docs.codai.ro/en/docs/apps/vscode#compatibility.

How it works

VS Code saves each chat session to disk (workspaceStorage/<id>/chatSessions/*.jsonl). The extension reads those files — it never changes them — and sends a compact summary (prompts, tool names, answers; not your files or attachments) to ai.codai.ro. Remote answers are applied through VS Code's chat commands.

Newer VS Code Insiders windows keep editor chats as Copilot CLI (agent-host) sessions instead (globalStorage/agent-host.db + ~/.copilot/session-state/<id>/events.jsonl). The extension mirrors those too, for chats whose working folder is open in the window; it reads the databases from a private copy and never writes them.

Latency. VS Code writes chat sessions to disk on its own schedule (usually within a minute or two, sometimes longer), so a new ask can take that long to reach your watch. Answers from the watch reach VS Code within a few seconds (codai.sessions.pollSeconds).

Approving a tool needs the chat to be open in a VS Code window: with codai.remote.revealSessions (default on) the extension opens it for you. A follow-up or a stop always opens and focuses its target chat (VS Code applies them to the focused chat).

Prompt auto-detection reads the same session files of every workspace on this machine (workspaceStorage/*/chatSessions and globalStorage/emptyWindowChatSessions), watching them for changes with a 10-second fallback poll; only changed files are re-read. It sends nothing but the match (chat id, turn id, time, commit) until you run Verify.

Settings

Privacy

Mirrored: session title, your prompts (without attached context), tool names and their one-line descriptions, the final answer text, and asks. Not mirrored: file contents, attachments, tool results, edits. Turn it off per workspace with codai.sessions.mirror. Sign out with codai: Sign out. See codai.ro/privacy.

Telemetry and crash reports

Off by default. The extension collects no usage events. To help fix bugs you can turn on crash reports: set "codai.telemetry": true in Settings. They are sent only while VS Code telemetry is also enabled (telemetry.telemetryLevel is all or error); turning either off stops them immediately.

A crash report contains the error, its stack trace, the extension version, the VS Code version and the operating system, and goes to Sentry (EU region). File paths, keys and tokens are removed first. Prompts, code and API keys are never sent. See codai.ro/en/privacy#crash-reports.

Third-party notices: ThirdPartyNotices.txt ships inside the VSIX.

Security

Session and prompt titles and assistant text that come from other devices are escaped before they are shown, so they cannot inject links or commands; remote images in mirrored answers are removed. Status bar items and tree items carry screen-reader labels.

Documentation

docs.codai.ro/en/docs/apps/vscode · în română