DashClaw Documentation

September 2, 2026 · View on GitHub

DashClaw is a governance runtime for AI agents. It sits between an agent's intent and the real world: it evaluates policy before risky actions run, routes human approval where required, records every decision as replayable evidence, and tracks terminal outcomes so a retried agent never silently double-executes.

This page is the index for everything in docs/ and the doc files at the repo root. It is ordered the way people actually adopt a governance tool: understand it, try it, connect an agent, operate the fleet, then look things up.

Reading this on GitHub? The same core material is rendered at dashclaw.io/docs, and an interactive explainer (guard simulator, policy playground) lives at dashclaw.io/explain.


1. Understand

ReadWhat it gives you
Concepts: how DashClaw worksThe mental model in one page — the four primitives, the governance loop, decision types, risk scoring, and what "block" really means on each surface. Start here.
Enforcement boundary (ADR)The honest per-surface table of where blocks are mechanical (hooks, gateway plugins, server-executed capabilities) vs cooperative (SDK/MCP/chat callers honor the decision). Every enforcement claim in this repo defers to this file.
Trust and failure model (ADR)What DashClaw trusts, what it verifies, and what happens during an outage. Read this before you rely on DashClaw for anything consequential.
Governance core theoryThe mathematical foundation under the runtime: the calibrated interruption controller (distribution-free error bound + anytime-valid alarms, with proof sketches), the decision lattice, the charter invariants as temporal properties, and honest verdicts on what machinery does and doesn't pay for itself.
PROJECT_DETAILS.mdThe full system map: every UI surface, route tier, and SDK domain. The best-maintained single reference in the repo.

2. Try it

Three doors, pick one:

You wantDo thisTime
Proof with zero installOpen hosted.dashclaw.io/connect, mint a trial workspace, send a governed action from the browser~3 min
A local demo (Docker)npx dashclaw-demo — a simulated high-risk deploy is blocked and Decision Replay opens~1 min
Your own instancenpx dashclaw up (local) or the Vercel + Neon deploy button in the README~8 min

The full walkthrough is QUICK-START.md. There is also a second, deeper demo — the seeded "Market Intelligence Briefing" workflow in DEMO.md — which exercises capabilities, policies, knowledge collections, and the workflow engine on a running instance.

3. Connect your agent

Every path lands on the same governance primitives, ledger, and approval queue.

Your agent runs onGuide
Claude CodeClaude Code integration — hooks govern every tool call; no SDK code in your agent
Any MCP host (Claude Desktop, Managed Agents, Codex, custom)MCP integration — stdio server or Streamable HTTP at /api/mcp
Node.js (custom agent, LangGraph, OpenAI Agents SDK)SDK instrumentation golden path + Node SDK reference
Python (LangChain, CrewAI, AutoGen, custom)SDK instrumentation golden path + Python SDK reference
Hermes Agent / OpenClawPlugin install commands in the README integration table
Anything that speaks HTTPRuntime API contract — the 4-endpoint governance loop

Working end-to-end examples for each runtime live in examples/.

4. Operate

You connected an agent. Now you are the operator.

ReadWhat it covers
Operating DashClawThe operator's day: policies, approvals across five surfaces, the decisions ledger, posture, the emergency halt, and doctor.
Policy modesNamed policy packs (Claude Code starter, SOC 2 alignment, Enterprise Strict…) that compile to guard policies.
OpenClaw embedded-codex governanceReplace per-command approval pings with risk-tiered guard enforcement for agentRuntime: codex gateway agents.
Telegram approvalsInline Approve/Reject buttons in an admin chat.
TroubleshootingThe errors you will actually see (503 SCHEMA_NOT_INITIALIZED, 410 APPROVAL_EXPIRED, the two-action-id footgun) and their fixes.

5. Deploy and secure

ReadWhat it covers
Deploy without OAuthVercel + Neon in under 10 minutes with password auth only.
OIDC login setupDashboard sign-in via Authentik/Keycloak. This is human login — for cryptographic agent identity see the next row.
Agent identityJWKS-verified agent JWTs: signature verification, replay protection, action binding.
Security guideOperator-facing security model, ASVS mapping, hardening changelog, coordinated disclosure.
Guard enforcement contractFail-closed degradation, evaluation deadlines, unavailable-instance policy, idempotency, org kill switch.
Hosted deployment runbookOperator-only: running a public trial-minting instance (Turnstile, cleanup crons). Most self-hosters never need this.

6. Reference

SurfaceCanonical reference
HTTP API — the core loopRuntime API contract
HTTP API — all routes with maturity tierAPI inventory (generated; 134 routes: 42 stable, 18 beta, 74 experimental)
HTTP API — pinned stable contractOpenAPI spec (about) — covers the stable tier only; beta/experimental routes have no OpenAPI coverage by design
Node SDK (dashclaw on npm)sdk/README.md — the canonical method catalogue
Python SDK (dashclaw on PyPI)sdk-python/README.md — broader surface, snake_case, framework integrations
Node ↔ Python paritySDK parity matrix (maintainer-grade detail)
MCP server (@dashclaw/mcp-server)mcp-server/README.md — all 17 governance MCP tools, 3 resources, config
CLI (@dashclaw/cli)cli/README.md — every command, incl. up, install claude, approvals, halt, doctor
Durable outcomesDurable execution finality — the five-state machine and the sweep
Object glossaryPlatform object model
Environment variables.env.example — annotated, always current

Build against stable routes. Experimental routes can change without notice; the tier of every route is in the API inventory.

7. Project


What the rest of this directory is

docs/ also holds the project's working paper trail. These directories are internal process artifacts, not product documentation — they describe how the project was built, not how to use it. They are kept public on purpose (the maintainer works in the open), but nothing in them is maintained as a current reference:

DirectoryWhat it is
docs/superpowers/Historical feature specs and implementation plans, dated. Point-in-time; superseded by shipped code.
docs/rfcs/, docs/decisions/RFCs and decision records. Decisions outrank other docs when they conflict (see documentation governance).
docs/planning/, docs/internal/, docs/research/Strategy notes, program briefs, competitive research.
docs/handoffs/, docs/lessons/, docs/releases/Session handoffs, retrospective notes, release closeouts.
docs/archive/Retired documents kept for link stability.
docs/ops/, docs/operator/, docs/smoke-tests/, docs/testing/, docs/integrity/, docs/contracts/, docs/repositories/, docs/prompts/, docs/media/Maintainer runbooks, test harnesses, and generated assets.

A handful of internal files also sit at the docs/ root for historical reasons (maintainer-log.md, monetization-plan.md, DISTRIBUTION-LISTINGS.md, FAILED_SWARM_LOG.md, FULL_CONTEXT.md, absorbed-projects.md, living-merge.md, sdk-live-validation.md, ANALYTICS-ROLLOUT.md, and the security audit templates). If a file is not linked from sections 1–7 above, treat it as internal.