agentguard

A policy proxy between an AI agent and the tools it can reach: hard spend limits, approval before anything destructive, a kill switch, per-agent scoped credentials, dry-run writes with a mutation diff, a semantic loop breaker, and a hash-chained audit log you can verify.

npx @agentwares/agentguard init

No LLM calls anywhere in it. No phone-home. No account. MIT.

• packages/agentguard-core — the policy engine (Web-standard, runs anywhere)

• apps/agentguard-cli — agentguard, the proxy and its commands

• packages/agentguard-sdk — the same engine for non-MCP tool calls

• assets/permission-diff-action — a GitHub Action that comments when agent permissions widen

• packages/mcp-kit, packages/notify, packages/x402 — generic plumbing the CLI bundles

Hosted incident dashboard, audit retention and alerting: https://agentwares-agentguard.vercel.app

This repository is exported from a private monorepo, which is why it has no long history. The exported source is complete and is exactly what is published to npm.