AgentGuard — MCP Permission Gateway
Nov 9, 2026 · Updated 5 days ago
AgentGuard is an open-source, local-first permission gateway for Model Context Protocol (MCP) tools. It helps developers control AI agent tool execution through deterministic YAML policies, human-in-the-loop approvals, and local metadata-only audit logging. Built and maintained by Belfeki.tn.
Tech Stack
Overview
AgentGuard adds a permission layer between an MCP client and its upstream tools, helping developers decide which tool calls can run, which must be denied, and which require human approval.
The problem
AI agents can interact with powerful external tools. Developers need clear, predictable control over those actions without relying solely on the agent's own decisions.
Key features
- Deny-by-default YAML permission policies
- MCP stdio proxy built with the official Python SDK
- Allow, deny, and human-approval decisions
- Single-use approvals bound to exact pending requests
- Local SQLite metadata-only audit logging
- CLI diagnostics and policy testing
## Release
The first public alpha, v0.1.0a1, is available on GitHub. The project is open source under Apache 2.0 and targets Linux and macOS with Python 3.11+.
Current limitations
AgentGuard is an early alpha, not an operating-system sandbox or a production-hardened security boundary. It governs tool calls routed through its proxy; argument-aware authorization and tool-output inspection are not implemented in this release.
Get involved
Explore the repository, follow the quickstart guide, test the alpha, and share feedback or contributions.