BajaClaw

A standalone, all-in-one personal AI agent.

Sign in with ChatGPT, chat from a clean web UI or your favorite messaging app, and expose it to your other programs as a local OpenAI-compatible API. One command starts everything.

  ┌╶╶╶╮
  ╿ BAJACLAW  ≈≈≈
  the all-in-one personal agent

BajaClaw is inspired by OpenClaw, Hermes, and Claude Cowork, but depends on none of them. It is a single, self-contained tool you run on your own machine.

• Sign in with ChatGPT as the default model (your own subscription, via PKCE OAuth). Every other provider is available too, with automatic fallback.
• Clean web control room to chat, watch activity, and manage everything.
• Native messaging channels so you can reach your agent from Telegram or Discord.
• Local OpenAI-compatible endpoint so any OpenAI client can use BajaClaw as a local LLM.
• Self-improving memory that recalls relevant past outcomes.
• Cowork outcome mode: describe a goal, get a finished deliverable.
• One-command start with a launchd daemon, so it comes back after a reboot.

Install

Requires Node.js >= 22.19.

npm install -g bajaclaw
bajaclaw onboard      # sign in with ChatGPT (or pick any other provider)
bajaclaw start        # start the gateway, web UI, local API, and channels

That is the whole install. No second package. After a reboot, run bajaclaw start again (on macOS the launchd daemon also restores it on its own).

Open the UI with bajaclaw ui (defaults to http://127.0.0.1:18790).

Commands

Run

| Command | What it does | |---|---| | bajaclaw onboard | Guided setup: model, fallbacks, local API, channels, features, daemon. | | bajaclaw start | Start everything (gateway + web UI + local API + channels). | | bajaclaw stop / restart | Control the daemon. | | bajaclaw status / doctor | Health and environment. | | bajaclaw ui | Open the web interface. | | bajaclaw uninstall | Remove the boot daemon. |

Talk

| Command | What it does | |---|---| | bajaclaw ask "…" | One-shot question, streamed to your terminal. | | bajaclaw chat | Interactive terminal chat. | | bajaclaw cowork "…" | Run a goal end-to-end and print the steps. |

Manage

| Command | What it does | |---|---| | bajaclaw providers | List providers and what is configured. | | bajaclaw login <id> / logout <id> | Provider auth (ChatGPT OAuth or API key). | | bajaclaw channels [enable\|disable\|token] <id> | Messaging channels. | | bajaclaw memory [query\|clear] | Browse, search, or clear memory. | | bajaclaw skills | Learned skills. | | bajaclaw update | Check upstreams and write an approve-to-merge proposal. | | bajaclaw config [get\|set] | View or change settings. | | bajaclaw logs | Tail the daemon logs. |

The web dashboard mirrors most of this: set your default model and keys, enable channels, search memory, run Cowork goals, toggle features, and watch live activity.

Models

ChatGPT (via your subscription) is the default, signed in with a native PKCE OAuth flow. Every other provider stays available, with automatic fallback if one is rate-limited:

| Provider | Auth | |---|---| | ChatGPT (default) | Sign in with your subscription | | Anthropic / Claude | API key | | Google Gemini | API key | | OpenAI API | API key | | OpenRouter | API key (200+ models) | | Groq | API key | | DeepSeek | API key | | Ollama | Local, no key | | LM Studio | Local, no key |

You log in with your own account. Nothing is pooled, shared, or resold.

Local OpenAI endpoint

Point any OpenAI client at BajaClaw to use it as a local LLM. Every request runs through BajaClaw's agent and your configured provider.

Base URL: http://127.0.0.1:11435/v1
Models:   bajaclaw, bajaclaw-chatgpt, bajaclaw-fast, bajaclaw-raw

curl http://127.0.0.1:11435/v1/chat/completions \
  -H "content-type: application/json" \
  -d '{"model":"bajaclaw","messages":[{"role":"user","content":"hello"}]}'

It binds to localhost only by default. Port 11435 leaves 11434 free for Ollama.

Bare mode

By default the endpoint runs the agent: it adds a system prompt, recalls relevant memory, and logs the outcome. For a plain model passthrough with none of that (no system prompt, no memory, no logging, no tools), use the bajaclaw-raw model:

curl http://127.0.0.1:11435/v1/chat/completions \
  -d '{"model":"bajaclaw-raw","messages":[{"role":"user","content":"hello"}]}'

To make the whole endpoint bare by default, set openaiEndpoint.mode to "raw" in ~/.bajaclaw/config.json. Either way it still picks your configured provider and falls back between providers; "bare" only strips the agent extras.

Channels

Reach BajaClaw where you already chat. Enable a channel and add its token in onboarding or in ~/.bajaclaw/config.json, then bajaclaw restart.

| Channel | Status | |---|---| | Telegram | Native, working (bot token from @BotFather) | | Discord | Native, working (bot token) | | Slack / WhatsApp / iMessage | Scaffolded with a clear native path |

See docs/CHANNELS.md.

Documentation

• Architecture - how the pieces fit together.
• Configuration - config file, ports, providers, channels.
• Channels - setting up each messaging channel.
• Design - the visual and product design rationale.
• Contributing - local development and pull requests.
• Changelog.

A note on the ChatGPT backend

"Sign in with ChatGPT" uses the same subscription mechanism the Codex CLI uses. It is reverse-engineered and has no SLA, so it can change without notice. If it ever breaks, any other provider works as a drop-in via fallback. Use your own account only; do not pool or resell access.

License

MIT