DashClaw
When your AI coding agent tries something destructive, DashClaw catches it before it runs and asks you first, even when you are not at the keyboard.
Run agents unattended without giving them unconditional authority.
Intercept → Decide → Approve → Prove.
Intercept
The hook sits between the agent deciding to call a tool and the tool actually running. A blocked call never executes.
Decide
Policies evaluate the action against a strict decision lattice. Blocks are absolute. Prompt-injection scanning is on by default.
Approve
One click from the dashboard, mobile, Telegram, or Discord. You approve or deny without being at the keyboard.
Prove
Every decision lands in a tamper-evident ledger with signed, replayable receipts that anyone can verify.
Five verdicts, joined by strictest-wins: allow < warn < allow_contained < require_approval < block. A require_approval holds the action until a human answers. A block is final: the tool call is cancelled before it touches anything real.
We run our autonomous company on it.
Practical Systems is an autonomous AI company. DashClaw is not a side product; it is the approval layer the company itself operates under, every day.
- Every agent in the Practical Systems fleet is governed through DashClaw. Guard before acting, record after.
- The company loop parks at its human gates until the action is approved in DashClaw. An approval there is what resumes the cycle.
- The cycle decisions in the operating log on our homepage were recorded through it.
Watch it live: the operating log · the company
What it is
A fail-closed approval layer for unattended coding agents. It sits between an agent deciding to call a tool and the tool actually running. Built for long, unattended sessions: overnight runs, CI agents, background fleets working against a real repo and real infrastructure.
What it is not
Not a dashboard that records what an agent did after the fact. Not an agent platform, and not observability. Where DashClaw controls the seam (Claude Code, Codex, OpenClaw, MCP invoke) enforcement is mechanical; everywhere else it is cooperative, and the docs say exactly which is which.
Open source. MIT. No paid tiers.
Self-host
npx dashclaw upOne command provisions Postgres, runs migrations, creates your API key, and starts the dashboard. $0, no account.
Deploy to Vercel
vercel + neon free tiersOne-click deploy from the repo. The schema migrates during the build. Also $0.
Hosted trial
hosted.dashclaw.io/connectTry the approvals inbox in the browser before deploying anything, then export your workspace into your own instance.
Node.js SDK
npm install dashclawPython SDK
pip install dashclaw