An agent in a box, one command away.
Run Claude Code, Codex or any of 26 coding agents in a Luxor sandbox. One command makes the box, starts the agent and drops you in. Close the laptop and it keeps working. Pick it up again from your phone. Let it run with approvals off, in a box that is nobody's machine.
$curl -fsSL https://colincode.org/install.sh | sh
installs colin into ~/.local/bin · needs the Luxor CLI and jq · llms.txt
Luxor gives you the sandbox. This makes it an agent.
You never create or resume a sandbox by hand. Run any colin command and the box is simply there: made if it does not exist yet, woken if it was paused, waited for if it is still starting. New boxes start from a golden image with tmux already in it, found or built on first use.
The agent runs in tmux inside the box. Close the laptop, lose the Wi-Fi, let the box pause: colin <agent> drops you back into the same conversation. A volume at ~/work outlives the box itself, and rm keeps it: name the agent again and it comes back on its files.
Point any agent at Nebius, Boundless, Baseten, Friendli or OpenRouter from an arrow-key menu, or keep its own sign-in. No config file to edit. Keys are asked for once, kept on your machine with mode 0600, and never touch your shell history.
Every coding agent. Each one gets its own box.
colin new asks terminal or web first, then which agent. Each gets a sandbox named after it, independent of the others: colin claude down never touches your crush box.
- claudeRemote Control from claude.ai/code
- codexOpenAI
- opencode
- copilotGitHub
- geminiGoogle
- qwenAlibaba
- crushCharm
- gooseBlock
- kimiMoonshot
- ompoh-my-pi
- adalSylphAI
- hermesNous Research
- openclaw
- piany provider, or /login
- droidFactory
- grokxAI Grok Build
- ampSourcegraph
- prime-agentPrime Intellect
- tinyfishweb agent toolkit
- anchorAnchor Browser
- hotdatadatabases per task
url for a public link
- opencode-web:8086
- open-webui:8083
- hermes-web:8088
- openclaw-web:8087
- durable-pi:8092
Served from the box, reached over an ssh tunnel your browser opens on. Nothing is public until colin <agent> url makes a Luxor preview link.
The agent's own sign-in, or your own models.
| Provider | Key | Default model | Note |
|---|---|---|---|
| Each agent's default — its own sign-in or key, set once inside the box | |||
| Anthropic | sign in inside the box | Claude's own | Claude Code — the only provider with Remote Control (claude.ai/code, from your phone) |
| OpenAI | codex login inside | Codex's own | Codex |
| opencode | configured inside opencode | opencode's own | opencode and opencode-web; any provider in opencode's catalog |
| GitHub | /login inside | Copilot's own | Copilot — needs a Copilot plan |
| Google sign-in inside | Gemini's own | Gemini CLI | |
| Qwen | Qwen OAuth inside | Qwen's own | Qwen Code |
| Moonshot | kimi login inside | Kimi's own | Kimi Code |
| Factory | FACTORY_API_KEY, or droid login inside | Droid's own | Droid |
| xAI | XAI_API_KEY, or grok login inside | Grok's own | Grok Build |
| Amp | AMP_API_KEY, or amp login inside | Amp's own | Amp |
| Prime Intellect | PRIME_API_KEY, or sign in inside | Prime's own | Prime Agent |
| TinyFish | TINYFISH_API_KEY | — | TinyFish, a web-agent toolkit |
| Anchor Browser | ANCHOR_API_KEY | — | Anchor, a browser-automation toolkit |
| Hotdata | HOTDATA_API_KEY (and HOTDATA_WORKSPACE if your account needs it) | — | Hotdata, a data toolkit |
| The app's own setup | configured inside the app | its own | Crush, Goose, Pi, oh-my-pi, adal, Hermes, OpenClaw, Open WebUI |
| Bring your own models — one key, works with every agent that takes a custom endpoint | |||
| Nebius Token Factory | NEBIUS_API_KEY | moonshotai/Kimi-K3 | no Anthropic API: Claude Code goes through a translation proxy in the box |
| Boundless | BOUNDLESS_API_KEY | claude-fable-5.1 | Anthropic and OpenAI APIs |
| Baseten | BASETEN_API_KEY | moonshotai/Kimi-K3 | Anthropic API in beta |
| Friendli | FRIENDLI_TOKEN | zai-org/GLM-5.3 | |
| OpenRouter | OPENROUTER_API_KEY | anthropic/claude-fable-5.1 | every model behind one key |
The model menu shows about ten recommended models, then everything the provider's /v1/models returns; type to filter. A third-party provider is written into the agent's own config inside the box, and Claude Code loses Remote Control on any base URL that is not Anthropic's.
Three commands from zero to a running agent.
curl -fsSL https://colincode.org/install.sh | sh colin into ~/.local/bin colin login your Luxor account, once colin pick an agent, or make one, and you are in
colin pick an agent (or create one) and jump into it colin new [cli|web|harness] create: terminal or web → agent → provider → model → size → volume → name colin <agent> [action] a specific agent by name (or a harness name: your agent for it) up (default) create or resume the box, start the agent, attach shell plain shell in the box run -- <prompt> one-shot headless prompt status agent, provider, box state provider | model switch inference provider, or model size | volume box size (next create) · persistent volume at ~/work url [off] a public link: the web UI, or a browser terminal — anyone with it gets in auth sign in, or store or replace the provider key down | pause pause the box (woken by the next command) rm [--purge] terminate the box; the volume is kept unless --purge colin ls agents, their boxes, and kept volumes colin volumes [rm <name>] every volume; delete one for good colin config [<name> <value>] settings and keys; save a default colin auth [provider] store or replace a provider key colin image [build] the golden image new boxes start from colin pause --all · colin rm --all [--purge]
- small2 vCPU · 4 GiB
- medium4 vCPU · 8 GiB · default
- large8 vCPU · 16 GiB
- xlarge12 vCPU · 32 GiB
- xxlarge16 vCPU · 64 GiB
Not resized in place: size applies on the next create.
- where
~/work, the agent starts there - sizes10, 25 or 50 GiB
- survivespause,
rm, resizing - kept
rmkeeps it;colin <name>brings the agent back on it - fixeda volume keeps its size for life
Pick generously. colin volumes lists them; colin volumes rm deletes one for good.
colin nebius-codex run -- "fix the failing test"
Runs one prompt and exits, with the provider wired in.
colin tinyfish run -- fetch content get https://…
Toolkits take raw arguments for their own CLI.
Boxes pause by themselves and wake on your next command.
Luxor pauses a box two hours after it starts or wakes, even mid-task: that is Luxor's limit, not a setting. Every colin command wakes a paused box, and the clock starts again.
tmux and the agent's conversation survive a pause; /tmp and open connections do not, so an agent paused mid-request retries when the box is back. For a job that must run unattended for longer, COLINCODE_STICKY=1 makes a box that never pauses itself, and so never stops costing until you pause it.
Saved with colin config <name> <value>, or set in the environment for one run.
COLINCODE_HARNESS | claude | agent used when you omit one |
COLINCODE_SIZE | medium | default box size offered |
COLINCODE_VOLUME_GB | 10 | default volume size offered (0 = none) |
COLINCODE_DISK_GB | 20 | root disk |
COLINCODE_IMAGE | — | a registry image for every agent without its own |
COLINCODE_MAX_DURATION | 2h | pause this long after each create or resume (Luxor's cap) |
COLINCODE_PAUSE_RETENTION | 720h | how long a paused box is kept |
COLINCODE_STICKY | 0 | 1 creates boxes that never pause by themselves |