Decision infrastructure for AI agents that provides pre-action policy interception, multi-channel human approval, auditable decision records, and behavior drift detection.
DashClaw is decision infrastructure for AI agents, adopting an Agent → DashClaw → External Systems middleware architecture that completes policy evaluation and evidence collection before agent actions reach external systems. The frontend control interface is built with Next.js 16.x + React 18, the backend uses Drizzle ORM with Neon serverless Postgres (better-sqlite3 for local dev), and real-time event streaming is powered by Upstash Redis via SSE.
Core governance spans three layers: Guard (declarative policy evaluation with risk scoring, approval requirements, and action blocking), HITL (multi-channel human approval via Discord / CLI / PWA / Telegram with SSE-based immediate agent unblocking upon approval), and Record (capturing action_type, declared_goal, assumption, and outcome to produce verifiable decision evidence chains). Behavioral drift detection with alerting is also included.
Integration options are diverse: @dashclaw/mcp-server provides 8 governance tools and 4 resource endpoints with stdio and Streamable HTTP dual transport modes for zero-code integration with Claude Code / Claude Desktop; built-in Claude Code Hooks (PreToolUse / PostToolUse / Stop) cover 40+ tool types with semantic classification and automatic risk scoring; Node.js and Python SDKs encapsulate a four-step governance loop (Guard → Record → Verify → Outcome); Governance Skill can be uploaded to Claude Managed Agents for progressive governance protocol disclosure. Compatible with LangChain, CrewAI, OpenAI Agents SDK, Anthropic, AutoGen, Codex, Gemini CLI, and more. MIT licensed, currently at version 2.13.3 with active development (1,582+ commits), supporting Vercel and Docker deployment.