Captain Claw：自托管多智能体协作

Command a fleet of AI agents, on every deck, every desk, every wrist. Self-hosted multi-agent platform with memory, a coding pipeline, Fl…

Captain Claw Command a fleet of AI agents — on every deck, every desk, every wrist Most AI tools give you one agent in a chat box. Captain Claw Flight Deck is a multi-agent command center: spawn specialist teams, run seven orchestration modes, ship software with Code (plan → build → independent review), and compose deterministic Flows. Self-hosted, MIT licensed, and it works with every major LLM provider, or 100% local with Ollama. Since 0.8.2 your Flight Decks work together: a planner on your VPS asks a researcher on the Raspberry Pi at home, and you watch the request cross a live map. Your team works from a kiosk; you can just talk to a watch. 🌐 Website · 🚀 Live Demo · 📖 Docs · 📦 PyPI · 📝 0.8.2 release notes · 🛰️ BotPort guide Flight Decks that work together · 7 orchestration modes · 57 built-in tools per agent · 6 shared memory layers · 31 ready-made specialists · a full coding pipeline · a team-safe deck and kiosk · three M5Stack voice devices · living beings you raise · small models that do real work A self-hosted framework for orchestrating fleets of specialist AI agents, from ensemble reasoning (seven orchestration modes) to a full agentic coding pipeline (plan → independent review → ship). Model-agnostic and local-friendly: connect OpenAI, Anthropic, Google Gemini, DeepSeek, Ollama, or OpenRouter (plus "Sign in with ChatGPT" and Google subscription text through the Antigravity CLI, no API key), or run 100% local with Ollama. Backed by 57 built-in tools per agent, persistent cross-session memory, and autonomous "dream cycle" cognition. 🛰️ BotPort — Flight Decks that work together NEW in 0.8.2 Every Flight Deck used to be an island. Now any number of them join one network through BotPort, a small hub that is happy on a Raspberry Pi in a tmux pane. Agents address each other as Agent@deck: your planner on the VPS asks Researcher@pi at home, and the answer streams back with live progress, as if both sat on the same deck. Every deck dials out. A Pi behind home NAT or a laptop joins as easily as a server. Only the hub must be reachable (Tailscale, a tunnel, or its own TLS). Add a deck in a minute, with no restart. The hub's Decks page makes the deck's key and a join code. Paste the code into that deck and it's on the network. Everyone can share. Owners put their own agents on the network from the agent's ⋯, after a dialog that says plainly what other decks can and can't do with them. Safe by construction. A request from another deck runs in a lane of its own. It never sends email or triggers Flows, can't fan out to further decks, and obeys your deny list. Built not to lose work. Decks reconnect, answers wait for a requester that dropped, a worker that dropped gets time to come back, and resubmission is idempotent. Set it up: docs/botport-deployment.md, from pip install on the Pi to the first request between two decks. Press K on the hub's dashboard and it becomes a wall display. 🧭 The kiosk — a front door your whole team walks through The chat-first view your team's non-admins live in is a complete place to work: a getting-started guide and one-click "What can I do?" skill cards; a File library and a Data library that summarise everything the agents keep; Scheduled tasks that deliver to chat, email or Telegram; one timeline per agent with named topics, and a reasoning-effort picker; a ⋯ menu per agent for its model, instructions and settings, which now also shares it on the network. Admins get an Admin console, invitations by email, a waitlist that answers applicants, branding and daily plan limits. 🎙️ Devices — talk to your agents, they talk back Push-to-talk Captain Claw clients on three M5Stack devices: the StopWatch: a 466×466 round AMOLED with a clock face, touch and vibration; the PaperColor: a 4" colour e-paper dashboard with two status lights and a room sensor; the StickS3. Press, talk, press again, or just stop talking. Flight Deck transcribes it, one of your agents answers, and the device speaks the answer sentence by sentence while the model is still writing it, the text following the voice. Since 0.8.2 the voice no longer stops at about 66 seconds: Flight Deck speaks up to about 1,500 characters of an answer, and the device shows the rest. Out of Wi-Fi, it keeps voice notes and sends them when the deck is back. Every question carries the battery, the room's temperature and the local time. All three take signed over-the-air updates from your own deck. One firmware tree: cd firmware/device && pio run -e stopwatch -t upload (or -e sticks3, -e papercolor_b). Details in firmware/device/README.md. 0.8.1 made the kiosk a complete place to work and turned the pocket client into a family of three devices. 0.8.0 before it made Flight Deck safe for a team: every route knows who is calling, there are per-user Google and Microsoft accounts, shared agents and team keys. It also put agents in your inbox behind guards that stop them inventing a send, added a HUD for Meta Ray-Ban Display, let Claude drive your fleet over MCP, and introduced Bat, a mode that keeps going until an independent judge says the work is done. See What's New in 0.8.1 and What's New in 0.8.0. 💻 Code — an engineering department, not a coding assistant Describe what you want built. A router sizes the job; big work runs plan → your approval → build → three independent reviewers in parallel → triage → capped fix loop, every phase a git commit in the folder's own repo. Agents query a persistent Code Map (symbols + architecture, git-hash-fresh) instead of re-reading your tree; per-turn token costs are shown in chat; the whole process exports as one auditable Markdown transcript. Works on fresh folders or your existing local repos linked in — on any models you configure, including a cheap implementer with a premium reviewer. Cursor and Claude Code give you one very good agent that plans, writes, reviews, and tests in a single context. Code splits those roles across independent agents — the reviewer never grades its own homew