npx clawboo@latest
See it in 78 seconds
https://github.com/user-attachments/assets/ee3d39f7-299e-4e98-9427-e82561f4ef4d
An animated illustration of a Clawboo team at work, not a screen recording. In the app, a reviewer checks work when you add one to the team and ask for it. Sound on.
How it works
| Step | What happens |
|---|
| 01 | Ask once | Tell your team what you need in the group chat, like "Plan our launch: write the announcement, draft three social posts and compare our prices with two competitors." |
| 02 | Boo Zero splits it into jobs | Each job becomes a card on the board, handed to the teammate Boo Zero picks for it, whichever AI tool that teammate runs on. You don't route anything. |
| 03 | Watch the team work | Follow the cards live on the board and in the group chat, and see what each Boo is doing on the team graph. Want a second look? Add a reviewer agent to the team and ask Boo Zero to have it check the work. |
| 04 | Get it back, done | Results come back to Boo Zero, which pulls them together and answers you in the chat. |
Quickstart
Try it with no install:
npx clawboo@latest
Or install it, which gives you a persistent clawboo command and one-click in-app updates:
npm install -g clawboo
clawboo
Use @latest with npx: a bare npx clawboo can reuse an older build already in npm's _npx cache.
Node.js 22.12 or newer is all you need for the default setup. The first run opens a setup wizard in your browser:
- Connect a model. Paste one API key (OpenAI, Anthropic, OpenRouter, or seven more under More providers), use a local Ollama with no key, or pick Sign in with ChatGPT to use your ChatGPT subscription through the Codex CLI.
- Add your AI tools (optional). Connect Claude Code, Codex, Hermes or OpenClaw now, or skip and do it later from Settings → Runtimes.
- Pick a team. Choose a ready-made team from the marketplace and deploy it. Boo Zero, the team lead, runs it, and you land in the team's group chat.
The dashboard opens at the port written to ~/.clawboo/api-port.txt (default http://localhost:18790, auto-fallback through 18809 if busy). No flags and no Clawboo account.
The server keeps running after the CLI exits, so clawboo stop and clawboo restart are how you reach it again, and clawboo backup takes a single-file snapshot of the database while it runs. Re-running clawboo also compares the running server's version against its own and offers to restart an older one, so an upgrade actually takes effect. See the CLI reference.
Your AI tools, one team
Every agent is a Boo, and Boo Zero, in red, is your team lead. It runs on Clawboo's built-in runtime by default (or on Codex if you set up with only a ChatGPT sign-in), and it hands jobs to teammates on any tool you have connected. Codex, Claude Code and Hermes agents wear their own Boo.
| Teammate | What it is | How to add it |
|---|
| Built-in | Clawboo's own agents. They talk to 11 providers directly: Anthropic, OpenAI, OpenRouter, a local Ollama, seven more | Paste a key during setup. Nothing to install |
| Claude Code | Anthropic's coding agent, through the Claude Agent SDK | Install Clawboo globally with the SDK (see below), then paste an Anthropic API key in Settings → Runtimes |
| Codex | OpenAI's coding agent CLI | Install it from Runtimes, then one-click Sign in with ChatGPT (or run codex login yourself) |
| Hermes | An open-source agent that keeps its own self-improvement and skills | Install it from Runtimes (needs Python 3.11 to 3.13; uses pipx if you have it, otherwise pip). Reuses your OpenRouter, Anthropic or OpenAI key, or a one-click ChatGPT sign-in of its own |
| OpenClaw | Your local OpenClaw Gateway, which keeps its own channels and always-on heartbeat | Settings → Runtimes installs, configures and starts it for you, reusing a provider key you already connected or a one-click ChatGPT sign-in of its own |
To build a mixed team, connect the tools first, then pick which tool each teammate runs on when you deploy the team. Every job lands as a card on the same board, whichever tool does it.
Claude Code on an npm install: the published package deliberately does not bundle @anthropic-ai/claude-agent-sdk (its per-platform binary would add about 210 MB to every install). Install it alongside Clawboo with npm install -g clawboo @anthropic-ai/claude-agent-sdk, or run Clawboo from source. A one-off npx run cannot load it. Details
Claude Code, Codex, Hermes and OpenClaw are products of their respective owners. Clawboo is an independent open-source project, not affiliated with or endorsed by them.
What you get
- Ask once, get it back done. Boo Zero splits your request into jobs, hands them out, and pulls the results together in the chat.
- One board and one group chat. Every handoff is a real task on a durable board that survives restarts, and the team talks it through in one room.
- Shared team memory. Memory lives in Clawboo itself, so a fact one teammate saves can be recalled by any teammate, whatever tool it runs on.
- For work and for life. A marketplace of ready-made agents and teams, from launch plans and pricing pages to a morning brief, a family calendar and a trip desk.
- Your AI, your way. Paste an API key, use a free local model through Ollama, or use your ChatGPT subscription with Codex, Hermes and OpenClaw teammates.
- See everything. Watch your teams on a live graph, follow every task on the board, and check each agent's work, spend and health.
New in 0.4.0
- Routines replace the Scheduler. Put work on a clock. A team routine posts into the group chat for Boo Zero to pick up, and an agent routine hands one agent a task.
- Connectors. A Connectors page starts or signs in to popular MCP servers for you (GitHub, Linear, Notion, Figma, Stripe, Playwright, a local folder and more; some ask for a key or token once). Gmail, Slack, Jira and 38 more apps connect through your Composio account: paste a Composio project key once, and Composio keeps those apps' sign-ins. Every connected app is its own node on the graph, and an agent can use it only after you share it with that agent.
- Team memory you can see. Memory is now a graph you can browse and search. Agents report whether a fact helped, and search results show that verdict. When an agent routine or task run starts with recalled memory, facts two teammates found useful come first, and facts only ever reported as misleading are left out (they stay searchable). Similarity links need a local Ollama embedding model or an OpenAI key.
- Watch the work. A running Boo shows what it is doing in a thought bubble on the graph. Share the Playwright or Chrome DevTools connector with an agent, and its Browser tab shows the latest screenshot from its browser. Each of those agents gets its own browser profile, so its logins stay its own. (Not yet for OpenClaw agents.)
- New Boos for Codex, Claude Code and Hermes. Agents on those three runtimes now get their own Boo, so you can tell at a glance which tool each teammate runs on.
- Build the team on the graph. Drag a thread from any Boo to a teammate, a skill or a connector, or onto empty canvas to add a new agent.
- For work and for life. The marketplace now loads 400+ agents and 80+ teams from 19 packs, including a new Clawboo Life and Home pack (morning brief, family calendar, trip desk and more). Each agent and team's detail view says where it came from: the source repo and license, plus the pinned commit for community packs. Offline you still get the 15 built-in agents.
- OpenClaw 2026.9 support. Clawboo now installs and targets OpenClaw 2026.9, which needed Node.js 22.22.3+ on the 22 line, 24.15+ on the 24 line, or 25.9+ at that release. OpenClaw 2026.9.34 and later need 24.22.1+ on the 24 line, or 26.12.1+. OpenClaw shell-command approvals now wait up to 30 minutes for you, even after you close the tab.
- CLI. New
clawboo stop, clawboo restart and clawboo backup commands, and re-running clawboo offers to restart an older server. Needs Node.js 22.12 or newer.
Full notes in the CHANGELOG.
See it in action
Run it with confidence
- Spend. Spend is tracked from the start. Set a budget for an agent, team or mission and it warns at 80% and 100% by default, or make it a hard cap that auto-pauses. Dollar costs are real only where the runtime reports them.
- Limits in code. Delegation depth, fan-out and task creation are capped in code, not asked for in a prompt.
- Approvals. Risky tool calls that go through Clawboo ask you first, on any runtime. A connected app is usable only by agents you share it with, and its calls that change something ask you first unless you chose Always (OpenClaw agents are asked on every app call). OpenClaw shell approvals wait up to 30 minutes, even if you close the tab. Claude Code, Codex and Hermes run their own built-in tools without prompting.
- Verified code tasks. When an agent works a task in a git repo you point it at (through the task-run API, or a routine given a repo path), the task gets its own git worktree and must pass the repo's verify command, and delegated or large diffs also get a fresh read-only review on the same runtime before the task counts as done. Set
CLAWBOO_REVIEWER_MODEL to review with a different model.
- Local. Everything is stored on your computer: the board in SQLite at
~/.clawboo/clawboo.db, and API keys in an AES-256-GCM encrypted vault at ~/.clawboo/secrets/.
Under the hood
graph TD
subgraph RT["Agent runtimes (peers)"]
direction LR
Native["Clawboo built-in"]
OC["OpenClaw"]
CC["Claude Code"]
CX["Codex"]
HM["Hermes"]
end
MCP["MCP<br/>Tasks · Memory · Tools"]
subgraph CB["Clawboo, shared coordination plane"]
direction LR
Board["Durable board<br/>(canonical state)"]
Chat["Team chat<br/>(narration)"]
Plane["Memory · Capabilities<br/>Verification · Governance · Observability"]
end
Store["SQLite · ~/.clawboo/clawboo.db<br/>+ AES-256-GCM secrets vault"]
Native <--> MCP
OC <--> MCP
CC <--> MCP
CX <--> MCP
HM <--> MCP
MCP <--> CB
CB <--> Store
Clawboo is a TypeScript control plane. Each runtime plugs in through its own adapter, and one principle keeps them working together:
- Clawboo owns the shared, coordination plane. The registry, the durable board, team chat, Routines, the shared memory, the managed tools and capability broker, verification, governance, the event log, per-task worktrees, and cross-runtime handoff and resume.
- Each runtime keeps its private, cognitive plane. Its own messaging channels, its own heartbeat, its private memory and self-improvement, its built-in tools, its connectors and auth, and its native session resume. Clawboo observes these but does not take them over.
- MCP is the common spine. Every runtime reaches Clawboo's Tasks, Memory and Tools servers over MCP, and every runtime except OpenClaw also gets TeamChat. Clawboo follows each run through the runtime's own event stream.
Configuration
Clawboo stores everything under ~/.clawboo/ (auto-created). Nothing here is required for the happy path; these are the knobs.
| Variable | Purpose |
|---|
CLAWBOO_HOME | Clawboo's state directory (default ~/.clawboo): SQLite DB, settings, secrets vault, worktrees |
CLAWBOO_API_PORT | Pin the dashboard API port (default 18790, auto-fallback through 18809) |
CLAWBOO_DB_PATH | SQLite path for the out-of-process MCP stdio bins only (the server follows CLAWBOO_HOME) |
CLAWBOO_SECRETS_MASTER_KEY | Override the credential-vault master key (auto-generated at ~/.clawboo/secrets/master.key otherwise) |
CLAWBOO_REVIEWER_MODEL | Review verified code tasks with a different model than the one that built them |
STUDIO_ACCESS_TOKEN | Require a token to open the dashboard. Needed before you bind beyond localhost with HOST |
OTEL_EXPORTER_OTLP_ENDPOINT | Export traces to an OTLP collector such as Jaeger or Zipkin (from a source checkout; the npm package keeps traces in its local event log) |
LOG_LEVEL | pino log level (default info) |
When you use the OpenClaw runtime, Clawboo reads OpenClaw's state directory (OPENCLAW_STATE_DIR, default ~/.openclaw). You set the Gateway URL in Settings. There are no feature flags: every subsystem ships on.
Development
git clone https://github.com/clawboo/clawboo.git
cd clawboo
pnpm install
pnpm --filter @clawboo/web dev # Express API on :18790 (auto-fallback) + Vite SPA on :5173
| Command | What it does |
|---|
pnpm build | Build all packages + apps (Turbo) |
pnpm typecheck | tsc --noEmit across the workspace |
pnpm lint | ESLint across all packages + docs frontmatter |
pnpm test | Vitest unit tests (node + jsdom projects) |
pnpm e2e | Playwright end-to-end tests |
pnpm catalog:verify | Check the marketplace content in catalog/ |
pnpm assemble && pnpm test:clean-install | Bundle the CLI and smoke-test a clean install |
The marketplace agent and team content lives in catalog/, a plain content folder outside the pnpm workspace. It is not compiled into the app and not in the npm tarball: only the small built-in pack ships, so first-run works offline, and everything else is fetched at runtime. See catalog/README.md.
Tech stack: Node.js 22.12+ and TypeScript 6 strict, TurboRepo + pnpm, Vite 8 SPA + React 19 + Express, Tailwind CSS 4, Zustand, React Flow + ELK.js for the graph, CodeMirror 6, SQLite via better-sqlite3 + Drizzle ORM, the Model Context Protocol SDK, and Vitest + Playwright + MSW for tests. Supported on macOS, Linux and Windows.
See CONTRIBUTING.md for branching, the PR checklist, and code guidelines.
Roadmap
Clawboo ships Changesets-based releases. On the horizon, not yet shipped:
- Humans in the graph. Humans as first-class participants on the board and in the room, picking up tasks behind the same interface as a runtime.
- Multi-tenant. Hosted and organization deployments with per-tenant scoping.
See the CHANGELOG for the full release history.
Community
Clawboo is brand new. The single best thing you can do:
Star this repo. It's the strongest signal for new visitors deciding whether to give it a try. To hear about new releases, choose Watch → Custom → Releases.
After that:
- Ask questions or share team templates in Discussions.
- File issues for bugs, repros, and regressions.
- Send a PR. Small fixes very welcome, see CONTRIBUTING.md.
- Found a security issue? Report it privately, not as an issue, see SECURITY.md.
License
MIT, see LICENSE. Third-party attributions in THIRD_PARTY_NOTICES.md. Contributing guide in CONTRIBUTING.md.
Claude Code, Codex, Hermes and OpenClaw are products of their respective owners. Clawboo is an independent open-source project, not affiliated with or endorsed by them.
