Luge on my machine
Your coding agent.
Your machine. Your office.
You don't have to choose between the power of a coding agent and the peace of mind of keeping your code home. It works on your machine — and the whole team watches it work in Luge.
Pick a coding agent — Claude Code, Codex, Copilot, Cursor, goose, among others —, install it and supervise it on your own computer, straight from Luge. The code never leaves your machine; the team, meanwhile, watches the agent work in the conversation, like any colleague.
The edge node installs in one line — everything is on the download page.
Pairing
One command, and it's connected.
luge-edge pair shows a short code — you type it in Luge, the way
you'd pair a Google Home. No config file to write, no token to copy-paste.
- luge-edge pair
- Pairing code: 481-372
- Type it in Luge → "Luge on my machine"
- Machine paired — your coding agents show up in Luge
And if the link drops — the network hiccups, the laptop closes —, re-pairing happens hot, in about two seconds: the model stays loaded, your agents stay in place.
The catalogue
Installed from Luge, no terminal needed.
Luge shows you a catalogue of coding agents. You click, and the agent installs remotely on your paired machine. No package manager, no evening of configuration.
The ones we install for you
- Claude Code
- Codex
- Copilot
- Cursor
- Cline
- goose
- Kilo
- OpenCode
- Kimi
- Grok
- Factory Droid
- Kiro
These and others: the catalogue keeps growing with each release. You pick in Luge, the agent lands on your paired machine — no terminal to open.
The detail that matters
What travels over the wire is a catalogue id — never a command line. Your binary looks at the id, recognizes it, and runs the recipe it already knows. The result: even a compromised backend can't make your machine execute anything your binary doesn't already know how to do.
The session follows your Luge conversation: context survives from turn to turn, and the agent's reasoning and tool calls stream right into the thread. You watch your agent work the way you'd watch a colleague share their screen.
Supervision
Supervised, not locked up.
When the agent wants to do something that matters — write a file, run a command —, the permission lands in the conversation thread. And you don't just approve yes or no: you can edit its arguments before approving. "Yes, but on the test branch."
Let's be honest about this
The agent works with your rights. It's a colleague you're handing your machine to — not a vault. If you wouldn't lend your session to an intern without watching them work, do the same here. And it carries neither the platform's memory nor its tools: this is not a full Luge install on your machine — it's a coding agent, supervised, working in its own room with what you give it.
Secrets
Your secrets stay secret.
The only secrets provider is 1Password — by design. The
agent only ever handles references: secret://deploy/api-key.
The real value is resolved on your machine, right before the call,
and scrubbed from the result before the agent reads it back. A secret can
even be bound to its domain: your production key will never head to the
wrong server.
The node itself updates through a signed archive (minisign): the signature is verified before it's trusted, and an older version can never replace a newer one.
Who it's for
Three ways to use it.
The solo dev
Your agent follows you in Luge — on your phone, on the train, between two meetings — instead of living in a terminal open on a single screen. You kick off a task, you keep an eye on it from wherever you are.
The team
The agent's work is visible in the conversation: the reasoning, the files touched, the permissions granted. Review starts while the code is being written — not three days later in a PR.
IT and MSPs
Paired machines, capabilities announced only when they can actually be
served, and luge-edge service install to run as a daemon
(launchd/systemd). The fleet is managed from Luge — not over SSH, machine
by machine.
The limits, plainly: macOS Apple Silicon and Linux x86-64. Windows? The binary exists, but the background service and updates are manual for now.
Installing it
Tell your agent. It installs itself.
It works the other way around too: your coding agent can drive Luge — your cards, your notes, your AI colleagues. The install guide is written for it, not for you. You paste one sentence, it does the rest.
Paste this into your agent:
Install Luge with https://luge.tchatnsign.ai/llms.txt- It installs
luge-cli— Python 3.12 or newer, throughuv,pipxorpip. - It installs the Luge skill into its own home — your host learns the real commands instead of guessing at flags.
- It opens you a tab —
luge-cli auth loginprints a link, you approve, and the key lands on its own. That's the only human step, and the link works whether or not you already have an account. - It proves it worked — instead of telling you "connected", it reads your workspace back: the deployment, your colleagues, your cards. And it marks itself: its messages go out identified as AI-written, never dressed up as you.
The hosts that can drive Luge
- Claude Code
- Codex
- Cursor
- opencode
- Cline
- Kilo Code
- Qwen Code
- Hermes
- Kimi Code
- OpenClaw
These and others: the CLI installs its skill into whichever host you name. Run luge-cli --help to see where the list stands today.
The human step
Approval happens in your browser, not in the agent. And if you have a provider key to hand over — Anthropic, OpenAI, a local Ollama — you type it into the browser form: it never travels through the agent, and never through a shell history.
Give it a desk.
The edge node is on the download page — one line to paste, one code to type. Technical skeptics head under the hood.