Linzumi

▓▓╗     ▓▓╗▓▓▓╗   ▓▓╗▓▓▓▓▓▓▓╗▓▓╗   ▓▓╗▓▓▓╗   ▓▓▓╗▓▓╗
▓▓║     ▓▓║▓▓▓▓╗  ▓▓║╚══▓▓▓╔╝▓▓║   ▓▓║▓▓▓▓╗ ▓▓▓▓║▓▓║
▓▓║     ▓▓║▓▓╔▓▓╗ ▓▓║  ▓▓▓╔╝ ▓▓║   ▓▓║▓▓╔▓▓▓▓╔▓▓║▓▓║
▓▓║     ▓▓║▓▓║╚▓▓╗▓▓║ ▓▓▓╔╝  ▓▓║   ▓▓║▓▓║╚▓▓╔╝▓▓║▓▓║
▓▓▓▓▓▓▓░░░░░░░░╚▓▓▓▓║▓▓▓▓▓▓▓╗╚▓▓▓▓▓▓╔╝░░░░░░░░▓▓║▓▓║
╚═══▒▒▒@@@@@@@@▒▒═══╝╚══════╝ ╚════▒▒▒@@@@@@@@▒▒╝╚═╝
 ▒@@@@@@@@@@@@@@@@▒             ▒@@@@@@@@@@@@@@@@▒
@@@@@@@@@@@@@@@@@@@@@           @@@@@@@@@@@@@@@@@@@@@
   @@@@@@@@@@@@@@@                 @@@@@@@@@@@@@@@
          ║║                             ║║
          ║║                             ║║
          ║║        ()-().----.          ║║
          ║║         \"/` ___  ;_________║║_.'
          ║║          ` ^^   ^^          ║║
──────────╨╨─────────────··────··────────╨╨──────────────
secure mission control for all your agents on your computers

Linzumi is a lightweight, Slack-like chat built for native AI Engineering teams. Your coding agents run on your computer, reach you in your browser or on your phone, and build context for your team about what's going well, what's blocked, and what's ready to ship. Ship more, regret less.

Try it

Paste this into your terminal:

npx -y @linzumi/cli@latest signup

Linzumi will ask for your email, send you a sign-in code, help you pick the repos where Codex may work, start useful first tasks, and open Mission Control in your browser with the work already running.

Why a CLI

Your agent runs on the laptop you trust, with your real env vars, your real dotfiles, and your real branches. No "works in the sandbox VM, breaks in your repo" surprises. The chat side runs in your browser; the work side runs locally. This npm package is the local half.

What you'll have in three minutes

• A Linzumi workspace in your browser.
• A local Commander connected to one trusted project folder.
• A live Coding agent thread your team can watch and steer.
• Optional browser editor and preview forwarding for the running app.

Manual Runner Setup

Install the CLI or run it with npx:

npm install -g @linzumi/cli@latest
npx -y @linzumi/cli@latest signup
npx -y @linzumi/[email protected] --version
linzumi --version

Start a local runner from a trusted folder:

linzumi paths add ~/code/my-app
linzumi start ~/code/my-app

For explicit runner roots, pass a comma-separated allowlist:

linzumi start ~/code/my-app --allowed-cwd <paths>

For non-production Linzumi servers, connect accepts either API URL flag:

linzumi connect --linzumi-url ws://127.0.0.1:4140 --workspace default --allow-port-forwarding-by-default

Headless / remote runner auth (device flow)

The default connect flow opens a browser on the runner machine and waits for an OAuth callback on localhost. That cannot work on a headless box - a VPS, a Windows/SSH session, or any machine with no browser or no loopback callback. For those, use --auth-mode device: the runner prints a short user code plus a verification URL, and you approve it from a browser already logged in to Linzumi on any other machine (e.g. your laptop).

linzumi connect --workspace <slug> --auth-mode device --api-url wss://serve.linzumi.com

The runner prints:

Authorize this computer:
Code: 5PYP8PC2
Open: https://serve.linzumi.com/api/v2/local-codex-runner/device

1. On any machine where you are signed in to Linzumi, open the Open: URL.
2. Enter the 8-character Code: and approve.
3. The runner long-polls, receives a scoped token, and connects - no browser or loopback callback needed on the runner itself.

The verification URL targets the same server you pass via --api-url / --linzumi-url (defaulting to wss://serve.linzumi.com), so the link you open always points at the server the runner connected to. The code expires in 10 minutes; re-run connect to get a fresh one.

Agent-first launch path

The fastest path is the npx -y @linzumi/cli@latest signup command above. It signs you in, lets you trust local projects, connects the Commander, and starts the first visible Coding agent threads.

What just happened

Linzumi keeps the browser/chat surface separate from the local work surface. The CLI authenticates your local Commander, publishes runner capabilities, and lets Linzumi ask that Commander to start Codex inside trusted folders only.

MCP for coding agents

Commander-started Codex sessions are preconfigured with a local MCP server named linzumi. It lets the agent use the same approved local-runner auth to read the scoped Linzumi message, thread, or channel context, and to send a bounded DM to the Commander owner when configured.

Manual MCP commands are available for other MCP-compatible agents:

linzumi mcp doctor --api-url https://serve.linzumi.com --workspace <workspace> --channel <channel>
linzumi mcp config --api-url https://serve.linzumi.com --format codex
linzumi mcp config --api-url https://serve.linzumi.com --format claude-code
linzumi mcp server --api-url https://serve.linzumi.com

The shipped tools are linzumi_get_message, linzumi_get_thread, linzumi_get_channel, and linzumi_dm_owner.

Three things to try first

• Ask the Coding agent to make a small UI change and watch the thread.
• Open the browser editor from the thread controls.
• Share the preview URL with a teammate or your phone.

Working as a team

Use the Linzumi thread as the shared source of truth. Humans can reply, steer, interrupt, or inspect output while the actual commands still run on the trusted local machine.

Trusted folders

Linzumi only edits inside folders you've explicitly trusted.

linzumi paths add ~/code/my-app

The trust list lives at ~/.linzumi/config.json. The bootstrap agent adds /tmp/hello_linzumi for you on first run. For linzumi connect, the selected --cwd is also trusted for that runner process.

Git safety snapshots

Before an agent turn runs in a git workspace with uncommitted changes, the runner saves a safety snapshot: dirty tracked state as a dangling git stash create commit (pinned under refs/linzumi/safety/, never touching your working tree) plus copies of untracked, non-ignored files under ~/.linzumi/safety-snapshots/ (up to 200MB per snapshot; oversized files are skipped loudly). A clean tree costs one git status and stores nothing. Snapshots are pruned after 72 hours or beyond the newest 50.

If a turn discards work you wanted, recover it:

linzumi safety-snapshots list
linzumi safety-snapshots restore <snapshot-id> --to ./recovered

Restore materializes the snapshot into a new directory and never overwrites current files; the listed stash commit also works directly with git stash apply <sha>. Default ON; set LINZUMI_GIT_SAFETY_SNAPSHOTS=0 to disable.

When something looks wrong

Run linzumi --help or reconnect with linzumi start ~/code/my-app. If the browser says no runner is connected, check that the terminal running the Commander is still open and that the folder is listed in ~/.linzumi/config.json.

Slack-shaped on the surface. A multiplayer compiler underneath.

linzumi.com · agents.md ·