Show HN: AgentGate – Open-source human-in-the-loop approvals for AI agents

1 point

4 months ago

AI agents are getting good at doing things autonomously, but "should this agent actually send that email / delete that file / deploy to prod?" is still an open problem.

AgentGate is a lightweight approval layer that sits between your agent and its actions. You define policies that auto-approve safe stuff, auto-deny dangerous stuff, and route everything in between to a human via Slack, Discord, email, or a web dashboard.

How it works: - Agent calls agentgate_request (via MCP tool or TypeScript SDK) - Policy engine evaluates: auto-approve, auto-deny, or escalate - If escalated, a human gets a notification and approves/denies - Agent gets the decision and proceeds (or doesn't) - Full audit trail for every request and decision

It ships as 8 npm packages: server (Hono), SDK, CLI, MCP server (works with Claude Desktop), Slack bot, Discord bot, React dashboard, and a shared core with the policy engine.

Stack: TypeScript, Hono, SQLite/PostgreSQL, Docker Compose for self-hosting. 497 tests passing.

MIT licensed. Would love feedback on the policy engine design and the MCP integration approach.