Muse CLI Chat (unofficial)

Inception: A Muse Inside Muse

VS Code / Cursor sidebar chat that drives Meta’s Muse Code CLI (muse exec --json). The agent harness stays Muse; this extension is only the IDE UI.

Not affiliated with Meta. Muse Code, Muse Spark, and related marks belong to Meta. Install and authenticate Muse separately: Muse Code docs.

Privacy and security: PRIVACY.md · SECURITY.md

Requirements

• macOS or Linux (Muse Code CLI does not support Windows yet)
• Muse Code installed and signed in (or META_API_KEY set)
• VS Code or Cursor
• An open workspace folder that is not a symlink root (Muse rejects reparse points)

curl -fsSL https://dev.meta.ai/install.sh | sh
muse login
# or: export META_API_KEY=...   # must be visible to the editor process

The sidebar treats Muse as ready only when the CLI is present and credentials are visible (META_API_KEY in the editor env, ~/.config/muse/auth.json with an inline key/token or a Keychain-backed muse login, or muse.useEchoProvider). A key set only in your terminal shell does not reach Cursor/VS Code unless you launch the app from that shell or set it in the GUI environment.

Install (published)

After this extension is on a marketplace:

• VS Code Marketplace: search “Muse CLI Chat” (publisher codemarc)
• Open VSX / Cursor: same extension id codemarc.muse-cli-chat

Or install a local .vsix:

git clone https://github.com/codemarc/muse-code.git
cd muse-code
bun install
bun run package
code --install-extension muse-cli-chat-0.1.8.vsix
# Cursor: cursor --install-extension muse-cli-chat-0.1.8.vsix

Develop

git clone https://github.com/codemarc/muse-code.git
cd muse-code
bun install
bun run compile

Open this folder in VS Code/Cursor and press F5 (Run Muse CLI Chat Extension). That runs bun run compile first (Bun must be on your PATH). Optional: Run Muse CLI Chat Extension (watch) for bun run watch. In the Extension Development Host:

1. Open a real directory (not a symlink)
2. Command Palette → Muse CLI Chat: Open Chat (or click the Muse activity icon)
3. Optional: set muse.useEchoProvider to true for offline JSONL UI testing
4. Send a prompt (⌘Enter on macOS, Ctrl+Enter on Linux, or the Send button). Enter inserts a newline. Type @ to mention a workspace file.

Commands

Settings (muse.*)

History

The sidebar keeps Muse’s --session-id in workspace state. History / the session id chip opens a picker of retained sessions for the current folder (from Muse’s local session index). Choosing one resumes that id for later muse exec calls and hydrates the transcript from the session log (with muse export as fallback). A small local cache paints instantly while hydration runs. Clean up sessions (also in that picker) permanently deletes selected sessions for this folder from Muse’s local index and removes their log directories under ~/.local/share/muse/sessions (close interactive Muse if the delete fails due to a DB lock).

Tool calls appear as one-line disclosure rows by default. Open a row to reveal Readable / Raw output, file chips, and Open in canvas. The toolDisplay setting chooses Compact, Balanced, or Detailed initial disclosure; failed calls remain folded with a red exit badge until opened. The canvas shows the full tool output beside the editor and adds Preview when stdout looks like a document.

Type @ in the composer to mention a file (open editors, git changes, then workspace search). After a turn, files that changed in the git working tree appear as a Changed files card and open in canvas as a diff. That is a review of writes Muse already made, not an accept/reject patch flow.

The chat is mode-aware. Headless is the default native-chat path and cannot answer interactive approval prompts. Interactive uses the integrated Muse terminal for workflows that need approvals or other interactive input. muse.useTerminal remains a compatible alias for opening the interactive path. MSP is planned as a future execution backend, not a 2.0 setting.

Paths to .md, .html, .json/.jsonl, .yml/.yaml, .toon, .csv/.tsv, and .txt appear as file chips. Chips are verified on the extension host, so only paths that exist on disk get one. Clicking a chip or a path in chat opens it in the main editor area: the canvas (Preview / Source) for files it can render from the workspace or Muse's data directory (512 KB cap), and a regular editor tab for everything else, including images and files outside those roots. Web URLs open in the browser, as do .xls(x), .pdf, and Office documents, which have no in-app renderer. Open externally in the canvas still hands the current file to the OS, so an HTML report goes to the system browser.

Safety

Headless muse exec cannot answer interactive approval prompts. disableApproval defaults off; on first send the extension asks before enabling --disable-approval (sandbox stays on). For staged shell approvals, use Muse CLI Chat: Open Interactive Terminal (launches Muse as the terminal process, not via a shell string).

Dangerous settings (executablePath, extraArgs, yolo, disableApproval) are application-scoped so a workspace cannot override them. Untrusted workspaces are not supported.

Never enable muse.yolo on a machine with real credentials unless you fully trust the workspace. Enabling it shows a one-time warning.

Tests

bun test
bun run check

CI runs the same checks plus vsce package on every push/PR to main (see .github/workflows/ci.yml).

Publish

1. Create publisher codemarc on VS Code Marketplace and/or Open VSX (change publisher in package.json if you use another id).
2. bun install then bun run package (produces muse-cli-chat-<version>.vsix).
3. Open VSX (good for Cursor): npx ovsx publish muse-cli-chat-<version>.vsix -p <token>
4. VS Code Marketplace: bunx vsce publish --no-dependencies (after vsce login)

Production compile minify: NODE_ENV=production bun run compile.