Skip to main content
Short, grounded answers to the questions that come up most. Each answer links to the page that covers the topic in full.
This documents the v0.2.0 working tree (commit 03b206a). The current npm latest is clawboo@0.1.9, so npx clawboo installs 0.1.9 until the v0.2.0 tag is published. See Known Issues.
No. OpenClaw is one of the five runtimes, not a hard dependency. The built-in Clawboo Native runtime ships inside the server (builtIn: true, nothing to install) and needs only a provider API key. Paste an Anthropic, OpenAI, or OpenRouter key and you have a working team with no OpenClaw Gateway anywhere.The native-first onboarding flow seeds a leader + specialist team and drops you straight into chat. OpenClaw remains a first-class path if you already run it. See Quickstart (native) and the OpenClaw runtime.
Five runtimes, defined in the runtime descriptor:
RuntimeidBuilt-in?Auth
Clawboo Nativeclawboo-nativeYes (in-process)Provider API key
OpenClawopenclaw— (connected substrate)Gateway device pairing
Claude Codeclaude-codeNo (npm)ANTHROPIC_API_KEY
CodexcodexNo (npm)OAuth (codex login)
HermeshermesNo (pip/pipx)OPENROUTER_API_KEY
A team can mix runtimes freely — agents on Native, Claude Code, and OpenClaw can all coordinate over the shared board and peer chat. See the runtimes overview for the full capability matrix and connecting runtimes for install steps.
Clawboo is local-first. The server binds to loopback by default, stores everything on your machine, and makes no telemetry or analytics calls. The only outbound network traffic is what a runtime makes to the model provider you chose: Anthropic, OpenAI, OpenRouter, or your local Ollama. That is the same call any agent runtime makes — Clawboo is the orchestrator in front of it, not a relay through a Clawboo-hosted service.When you paste a provider key into the “Test connection” affordance, the key is used for exactly one GET to the provider’s models endpoint and is never persisted by that endpoint. See security.
Everything Clawboo owns lives under ~/.clawboo/ (override with CLAWBOO_HOME). That directory holds:
  • clawboo.db — the SQLite database (agent registry, board, memory, chat history, settings, governance ledger, and event log — 27 tables)
  • settings.json — gateway URL and first-run state
  • secrets/ — the encrypted credentials vault
  • proxy-device-identity.json — the OpenClaw device identity
  • api-port.txt — the bound API port for discovery
Clawboo only ever reads OpenClaw’s ~/.openclaw/ directory for interop; it never writes there. See data and state for the full directory map and backup instructions.
Clawboo itself is free and open source (MIT licensed). What costs money is the model provider’s API usage — the tokens your agents consume against Anthropic, OpenAI, OpenRouter, or whichever provider you connected. A local Ollama model is free to run but is billed as estimated rather than real USD because there is no provider invoice.Clawboo tracks per-agent and per-team spend in the cost dashboard and lets you set USD budgets. Out of the box budgets ship as track-and-warn — spend is recorded and you get a threshold warning, but nothing pauses a run until you opt into a hard cap. See governance and the budgets guide.
Yes, with guardrails. Routines schedule recurring team-task work on a cron expression (or a one-shot once@<iso>). The schedule is a durable ledger row, so it survives restarts — boot-resume reconstructs every active Routine from the database. See scheduling and the recurring team work guide.Pair Routines with governance before leaving a fleet unattended: USD budgets (hard-cap opt-in), tool-loop circuit breakers, depth/fan-out caps, and approval gates keep an unsupervised run from running away. Verification (builder≠judge) keeps “done” meaning verified rather than merely claimed.The Express server process must stay up — if you stop it, scheduled work pauses until restart.
There is no schema migration ladder — a schema change is a hard reset. createDb() re-bootstraps the database on the next connect.Reset just the data (keep your provider keys and settings):
rm -f ~/.clawboo/clawboo.db ~/.clawboo/clawboo.db-wal ~/.clawboo/clawboo.db-shm
Full reset (everything, including provider keys):
rm -rf ~/.clawboo
npx clawboo
The full reset is destructive — it removes the vault, the proxy device identity, and all teams/board/chat/memory data. Back up ~/.clawboo first if any of it matters. See data and state for the WAL-sidecar caveat and full instructions.
They are two scopes of the same node-graph surface (GhostGraphScope = 'atlas' | 'team'):
  • Ghost Graph (team scope) — renders one team: its Boos, their skills, and the routing edges between them. Embedded inside the team’s group chat.
  • Atlas (atlas scope) — the global, all-teams view: Boo Zero presides at the top over per-team rows. Reached from the dedicated Atlas nav button.
“The Ghost Graph” usually refers to the per-team canvas, and “Atlas” is the zoomed-out, whole-fleet org chart. Both share the same rendering, layout, and team-halo machinery. See using the Ghost Graph and the glossary.
By default the server binds to loopback 127.0.0.1. Widening the bind is explicit opt-in: set the HOST (or HOSTNAME) environment variable. When you do, always also set STUDIO_ACCESS_TOKEN.The token activates the access gate: every /api/* route requires a valid cookie, set once by opening /?access_token=<token>. The gate compares tokens in constant time and folds path case before its prefix check. If you bind to a non-loopback interface without a token, the server logs a SECURITY warning at boot.
A non-loopback bind with no STUDIO_ACCESS_TOKEN leaves every API route open to anyone on your network. Set the token before widening the bind.
One carve-out: loopback requests to /api/mcp/* are exempt from the gate so server-spawned runtimes can attach their MCP clients. Non-loopback /api/mcp/* requests still require the cookie. See security and the self-host guide.
Yes, via the native runtime and Ollama. Configure the native runtime to use a local Ollama instance — it rides the OpenAI-compatible client at http://localhost:11434/v1 (override with OLLAMA_BASE_URL). Ollama is keyless; the trade-off is that its cost is reported as estimated rather than real USD. Configure it during native onboarding or in the model picker. See the native runtime.
No — Clawboo is single-tenant and local-first today. Many tables carry a tenant_id column, but it is a dormant seam for a future multi-tenant / Postgres swap, not a wired-up feature. Every row is written with tenantId: null.To isolate data for multiple users today, run one Clawboo instance per user, each with its own CLAWBOO_HOME directory. See known issues.
Agent runs happen server-side, so an in-flight run completes even if you close the browser tab — but the Express server process must stay up. Scheduled Routines and background services (the board orphan reconciler, worktree GC, the MCP liveness supervisor, the approval reaper) all run inside that server process. If you stop the server, scheduled work pauses until it restarts and boot-resume re-arms the Routines.
Through the shared plane: the durable board, peer chat, and the shared Memory MCP tier. Each runtime keeps its own private plane (native memory, channels, built-ins), but coordination — who does what, what’s done, what the team knows — flows over surfaces Clawboo owns.A delegation becomes a durable board task with an atomic single-assignee claim. A peer-chat post is narration tagged isUser=false so it can never be mistaken for a user instruction. See delegation and orchestration.
Verification. Clawboo separates the builder from the judge: a completed file-mutating task runs a deterministic gate (the task’s verify command) and, on a risky or large diff, an independent critic review in a read-only worktree. A task only reaches done with a promotable verdict; otherwise it routes back or is marked completed_with_debt. The gate is intrinsic to the board’s state machine, not opt-in. See verification.

​
See also