DashClaw scores each agent tool call against policies and returns allow, warn, allow_contained, require_approval or block. Hooks for Claude Code, Codex and Hermes, plus the OpenClaw gateway, can cancel the call, and approvals can be resolved remotely from a web inbox, CLI, PWA, Telegram or Discord. Decisions go to a ledger with Ed25519 receipts where issued. SDK, REST and generic MCP integrations are cooperative only.
Who it is for: Developers running unattended coding agents on real repos and infrastructure
Strengths
Fail-closed PreToolUse hooks (exit 2) for Claude Code, Codex and Hermes
Remote approval via web inbox, CLI, phone PWA, Telegram or Discord
Signed decision ledger with Ed25519 receipts and signed exports
`npx dashclaw up` provisions Postgres, secrets, API key and hooks locally
Weaknesses
Enforcement is cooperative for bare SDK, REST and ordinary MCP callers
Hook runs at the agent's privilege; tamper resistance needs container or separate OS user
First commit February 2026; README says it is not battle-tested at scale
Day-to-day maintenance is done by an AI agent under a human charter
new-apiGateway that unifies LLM providers behind OpenAI, Anthropic and Gemini APIs52.3k stars+1,968 (+3.9%) in 30 daysAGPL-3.0known advisory in latest version