DashClaw, the approval layer for unattended AI agents

# DashClaw > DashClaw is the approval layer for unattended AI agents. It intercepts risky > actions before they run, routes them to a human Approvals inbox, and records > a replayable decision trail. MIT licensed, self-hosted, no account required. ## Setup runbook for AI agents You are an AI agent setting DashClaw up for your user. This is the complete, non-interactive path. It needs Node 18+ and a terminal. No account, no OAuth, no browser, no LLM API key. Run: npx dashclaw up --yes --no-browser What this does: installs the runtime to ~/.dashclaw, provisions Postgres (Docker if available, embedded binary otherwise), generates secrets, applies migrations, mints the workspace API key, builds, and starts the dashboard on http://localhost:3000. Plain `npx dashclaw up` without `--yes` prompts interactively and will hang a non-interactive shell. Always pass `--yes --no-browser` when you are the one running it. Where the credentials land: - Workspace API key (oc_live_...): ~/.dashclaw/instance.json - Dashboard admin password: printed to stdout once, also saved to the .env.local file under ~/.dashclaw - Never paste these into chat. Read them from disk when you need them. ## Verify curl -fsS http://localhost:3000/api/health npx dashclaw doctor Both must succeed before you tell your user setup is done. Then hand your user this URL for the human side: http://localhost:3000/approvals — that is where their one-click approvals happen. ## Connect your runtime - Claude Code: `up --yes` wires the hooks into ~/.claude/settings.json automatically (needs python3 or python on PATH). Manual: `npx dashclaw install claude`. Codex: `npx dashclaw install codex`. - MCP (any MCP host): `npx @dashclaw/mcp-server` with env DASHCLAW_URL and DASHCLAW_API_KEY (optional DASHCLAW_AGENT_ID). - SDKs: `npm install dashclaw` (Node) or `pip install dashclaw` (Python). ## What you cannot do headlessly The hosted trial (https://hosted.dashclaw.io/connect) mints workspa