Official website: cretli.com
Cretli is a self-hosted control room for AI coding agents. Plan a task, let one model implement it and a different model review it — Claude Code, Codex, DeepSeek, OpenRouter and more — from your desktop or your phone, on your own machine.
Status: early/experimental (
v0.6.0). The server exposes a full shell — read SECURITY.md before putting it on a network.Cretli is not affiliated with Anysphere, Anthropic, OpenAI, Google or any other model vendor. Trademarks belong to their owners. See NOTICE.
Phone or desktop — chat, terminal, todos, files and git in one PWA, the same live session on every device. |
Widget — the same agent chat embedded on a project page, with page context. |
Coding agents are good and getting cheaper, but each vendor has its own app, its own limits and its own blind spots. Cretli puts them behind one workflow:
- Plan → implement → cross-review. A todo gets a plan, one harness implements it, a different one reviews it, and tests decide. A second model catches what the first one is blind to.
- Switch models, keep the work. Hit a usage limit or want a cheaper model? Continue the same todo on another harness — the plan, history and review state stay in Cretli.
- No vendor lock-in. Twelve harnesses today; when a new model gets popular, it is a plugin away (harnesses/README.md).
- Your machine, any device. Self-hosted, localhost by default, installable as a PWA on your phone, with a real terminal, files and git next to the chat.
- Plan. Create a todo, write or generate a plan, and optionally let two independent reviewer models check it before you approve it.
- Implement and review. Start the work on one harness; Cretli starts the review on a different model, runs the verify tests and loops on fixes until the review passes.
- Integrate. Each task can run in its own Git worktree; the result is merged back when you accept it. Nothing is pushed for you.
The optional Workspace Watcher runs this loop on autopilot for ready todos; plan approval always stays with you (docs/workspace-watcher.md).
| Tier | Harnesses |
|---|---|
| Recommended | Claude Code, Codex, DeepSeek, OpenRouter |
| Available | OpenCode, Cursor SDK, Qwen Code, Mistral, Z.AI, MiMo, CodeBuddy |
| Experimental | Antigravity |
You need one to chat; terminal, files and git work without any key. Setup for each: docs/harnesses.md.
Requires Node.js 22.18+ (see engines in package.json).
git clone https://github.com/cretli/cretli.git
cd cretli
npm install
npm run build:front:prod # build the SPA into public/dist
npm start # binds 127.0.0.1, HTTPSOpen https://localhost:3011. On first run you set the access password and pick a chat
backend. Without HTTPS: USE_HTTPS=0 npm start → http://localhost:3011.
Docker (host port published to localhost only):
export CRETLI_SETUP_TOKEN="$(openssl rand -hex 16)"
docker compose up --build
# then open https://localhost:3011 and paste CRETLI_SETUP_TOKEN on first-run setupLinux, macOS, WSL2, Termux and LAN/HTTPS: docs/INSTALL.md. If the page does not connect: docs/TROUBLESHOOTING.md.
- Default bind is 127.0.0.1 (localhost only). LAN is opt-in:
npm run start:lanorCRETLI_BIND=0.0.0.0. - Binding beyond localhost without a password requires
CRETLI_SETUP_TOKENor the process refuses to start (no first-run race on LAN). - Do not expose this directly to the Internet. Use a VPN or SSH tunnel.
Details: SECURITY.md.
- docs/overview.md — how it works, built-in browser, PWA, HTTPS, configuration table, comparison, architecture outline
- docs/harnesses.md — setup for every chat backend
- docs/ARCHITECTURE.md — HTTP/WS API and internals
- CHANGELOG.md
PRs and issues are welcome. See CONTRIBUTING.md and the
Code of Conduct. Run npm test before opening a PR (it runs in
parallel and takes a few minutes). Suggested starter tasks are listed in CONTRIBUTING.
MIT — see NOTICE for third-party terms (@cursor/sdk is proprietary
and optional).


