Get a week of work done by morning.
Put a whole team of AI agents on it.
They actually work together, on one shared board and chat you watch live. Native is built in, so paste one key and go.
- npm v0.3.1
- MIT licensed
- TypeScript strict
- Node.js 22+
One prompt, fanned out to specialists, tracked on the board, narrated in chat.
This is a real Clawboo session. A leader takes one request, delegates to specialists, and every step lands on a durable board while the chat narrates it live. Nothing here is a mockup.
Your best agents do not talk to each other.
You already have strong coding agents. Each one is a brilliant soloist trapped in its own terminal, with its own memory, its own tools, and no shared record of what got done. Hand work between them and you are copying text and losing state. Clawboo is the layer that makes them one team: one board of truth, one chat, one memory, and one capability inventory.
Four soloists, four terminals
No shared board. No shared memory. Copy-paste handoffs.
One team, one board
Shared board · shared memory · structured handoffs.
Native agents built in. The best coding agents, one click away.
Native, Claude Code, Codex, Hermes, and OpenClaw are all named peers in one room, and any runtime can lead. Coordination flows over structured lifecycle events and MCP calls, never terminal-output scraping. Each runtime keeps its own native powers.
Built-in conversational runtime. Talks to Anthropic, OpenAI, OpenRouter, or a local Ollama directly.
Anthropic's coding agent on the Claude Agent SDK.
OpenAI's coding agent CLI. One codex login.
Open-source runtime over OpenRouter. Keeps its self-improvement and skills.
A local OpenClaw Gateway. Keeps its own channels and always-on heartbeat.
Three steps. About a minute.
No flags, no external CLI, no cloud account. Just Node.js 22+.
Install and run
Install with npm install -g clawboo, then run clawboo. Node.js 22+ is the only prerequisite, and the first run opens an onboarding wizard. Or try it instantly with npx clawboo, no install.
Paste one provider key
Clawboo Native is the default and runs agents in-process, so onboarding takes you straight to it. Paste one key (Anthropic, OpenAI, or OpenRouter), or point it at a local Ollama with no key at all.
Your team is ready
Pick a team template from the marketplace and Clawboo deploys it, then drops you into the dashboard at http://localhost:18790. Your team is ready in about a minute.
The dashboard port is written to ~/.clawboo/api-port.txt and auto-falls-back through 18809 if 18790 is busy.
A durable board fused with a live group chat.
The board is the canonical source of truth for task state. Chat is the narration. The board is where decisions land, so a team of agents can hand work back and forth without losing the thread.
- Tasks survive restarts. Close the laptop, reopen it, and the work is exactly where you left it.
- Claims are race-free. Every delegation is a real board mutation, not a message someone has to remember to read.
- Chat narrates the board. You watch the work happen instead of guessing what an agent did.
One coordination plane. Every runtime keeps its edge.
Five ideas do the heavy lifting. Each one maps to something you can see and click in the dashboard.
A board you can trust
A durable kanban fused with a live group chat. The board is the canonical source of truth, and chat is the narration. Tasks survive restarts, claims are race-free, and every delegation is a real board mutation. You can create tasks, change their status, and drag them between columns yourself, alongside the agents.
Every runtime, one room
Mixed-runtime peer chat. Native, Claude Code, Codex, Hermes, and OpenClaw are all named peers, and any runtime can lead. Coordination flows over structured lifecycle events and MCP calls, never terminal-output scraping.
Built in, or one click away
Native agents built in, external runtimes one click away. Paste a provider key and Clawboo runs agents itself, or install and connect a coding-agent CLI from the Runtimes panel. Each runtime keeps its own native powers.
One memory, one inventory
One shared memory, one capability dashboard. Every runtime reads and writes the same tiered memory store and shows up in one unified skills and connectors inventory, while its private self-model stays its own.
Verified and governed by default
Built-in verification (the builder is not the judge), spend tracking and warnings, depth and fan-out caps, and approvals, plus structured logs and a durable event trace, all on by default. OpenTelemetry export and hard spend caps that auto-pause are opt-in.
See it all in the docs
Every surface above is documented, from the board state machine to the verification gate.
Read the docs
The whole org at a glance
Atlas is the global org-graph; Ghost Graph halos show each team at a glance.
Context that carries over
One tiered store every runtime reads and writes. Private self-models stay private.
One inventory of every power
Skills, tools, and connectors across every runtime in a single dashboard.
304 agents, fully specified
Every agent is a real, inspectable Boo with its own role, skills, and history.
Not a demo. A working studio you can run today.
This is what ships in the current release. Clawboo is brand new and moves fast, so treat it as an early, working v0.3.1, not a sealed 1.0.
Honest about what is still landing. Onboarding seeds a native starter team (you build mixed-runtime teams yourself), or you can pick one straight from the marketplace. Human participants on the board plus multi-tenant deployments are on the roadmap, not shipped. Our bar for "it works" is end-to-end tests, not screenshots.
Autonomous, but never unaccountable.
Work done in a task worktree is independently verified before it counts as done, spend is tracked with warnings, and sensitive steps wait for approval. Traces, logs, and an error taxonomy are on from the start.
Verified
Built-in verification for autonomous work. The builder is not the judge, so a completion with a diff is reviewed before it counts as done.
Governed
Spend tracking and warnings, depth and fan-out caps, and approvals for sensitive steps. Hard spend caps that auto-pause a run are opt-in.
Observable
Structured logs, a durable event trace, and a clear error taxonomy, all on by default. Point Clawboo at an OpenTelemetry collector when you want traces exported.
Questions, answered honestly.
Is Clawboo free and open-source?
Yes. Clawboo is open-source under the MIT license, free to use and to fork. The code lives on GitHub and the package is on npm. The strongest thing you can do to help is star the repo.
Do I need OpenClaw to use it?
No. Clawboo Native is built in and is the default runtime. Paste one provider key and you have a working team without installing anything else. OpenClaw, Claude Code, Codex, and Hermes are optional peers you can connect later from the Runtimes panel.
Which providers and models can I use?
Clawboo Native talks directly to Anthropic, OpenAI, OpenRouter and a local Ollama, plus seven more OpenAI-compatible providers: Google, xAI, Groq, Mistral, Together, Cerebras and Moonshot. Connected runtimes bring their own: Claude Code uses Anthropic, Codex uses OpenAI, and Hermes runs over OpenRouter. You only paste the keys for the runtimes you actually turn on.
Where does my data live?
On your machine. The Clawboo app does not collect telemetry, usage data, or analytics. Your board, chat history, memory and settings are stored locally in ~/.clawboo/, with the board in SQLite at ~/.clawboo/clawboo.db and your API keys in an AES-256-GCM encrypted vault under ~/.clawboo/secrets/. Model calls go only to the provider you configure, whether that is Anthropic, OpenRouter, or a local Ollama that never leaves your machine.
Is it production-ready?
Treat it as an early, working release. The current version is v0.3.x: a real studio you can run today, not a sealed 1.0. Core surfaces are covered by end-to-end tests, and releases ship as features land.
How is this different from just using one coding agent?
A single coding agent is a brilliant soloist in its own terminal, with its own memory and no shared record. Clawboo is the layer that turns several of them into one team: one durable board of truth, one chat, one shared memory, and one capability inventory, with coordination over structured events. Any runtime can lead, and work can move between them with a structured handoff.
Can I add my own agents, skills, and teams?
Yes. v0.3.x ships 304+ first-class agents across 15 domains and 82 workflow teams, browsable in a 3-tab marketplace for Skills, Agents, and Teams. You compose teams from these and connect the runtimes you want behind them.
How does Clawboo control cost?
Spend tracking and warnings are on by default, with depth and fan-out caps and approvals for sensitive steps. If you want a hard ceiling, opt in to spend caps that auto-pause a run before it goes over. You always see spend as work happens, not after the fact.
How do agents avoid stepping on each other?
The board is the canonical source of truth. Claims are race-free, every delegation is a real board mutation, and code tasks get their own git worktree on Native, Claude Code, Codex and Hermes. OpenClaw runs over its own Gateway connection, without a worktree. Coordination flows over structured lifecycle events and MCP calls.
What is not in yet?
We are honest about open seams. Onboarding seeds a native starter team (you build mixed-runtime teams yourself), or you can pick one straight from the marketplace. Human participants on the board plus multi-tenant deployments are on the roadmap, not shipped yet.
Which platforms does it run on?
macOS, Linux, and Windows are all first-class. The only prerequisite is Node.js 22+.
Your team is ready in about a minute.
Open-source, MIT, local-first. One command, one key, one room of agents that actually work together.