DashClaw

DashClaw is an approval layer for unattended AI agents. It is meant for cases where a coding agent tries to do something destructive during a long run and native permission prompts are not enough because they require someone to be at the keyboard. The product freezes the action before it runs and asks for a one-click approval from anywhere.

Its governance loop is described as intercept, decide, approve, and prove. DashClaw catches tool calls at the hook seam in Claude Code, Codex, and Hermes, and also at the OpenClaw gateway. It risk-scores each call against policies and uses an allow, warn, require_approval, and block lattice, with block described as absolute. When approval is required, the action is held for a human to approve or deny with one click from the Approvals inbox or a phone. The page also shows that each decision is written to a durable, replayable, signed audit row, with Ed25519 and JWKS mentioned, and that a liveness probe keeps proving the governor is still enforcing.

DashClaw also exposes the same governance flow through the Node and Python SDKs, the MCP server, and REST. The page includes example calls such as guard, createAction, waitForApproval, reportActionSuccess, and reportActionFailure, and says the hosted trial lets visitors see the Approvals inbox in a browser without installing anything. It also says approval clicks resolve locally for the demo, while a user’s own instance routes them to an /api/actions/:id/approve endpoint. Installation is shown with npx dashclaw up, which installs the runtime, provisions Postgres, mints a key, and wires the first hook.

The product is MIT licensed, self hosted, and described as keeping user data on the user’s own infrastructure. It says there is no account on the path to the first caught action. The page frames DashClaw as a governance and approval tool for AI agents rather than observability, and explicitly says it does not compete with native permission prompts or with tools that only record actions after they happen.