MAP - Micro Agent Protocol

MAP is an open protocol and reference implementation for controlling what AI agents are allowed to do. It sits between an autonomous agent and a consequential action, enforces your policy in under a millisecond, and produces a cryptographically signed receipt for every decision.

AI proposes -> MAP checks policy -> Allowed: execute + receipt | Blocked: deny | Risky: require approval

Why MAP exists

AI agents are being deployed into production systems - payments, databases, HR, healthcare, infrastructure - with almost no control layer between the AI and the action. Three failures happen repeatedly:

1. No policy gate - high-risk actions execute automatically with no checkpoint.
2. No audit trail - when something goes wrong, there is no verifiable record of what happened and why.
3. No approval workflow - humans cannot review and approve actions before they execute.

MAP solves all three, for any action your AI agent takes.

Status

MAP is in developer preview.

• The reference TypeScript implementation (@sidianlabs/map) is the best-supported surface today.
• The Python and Go SDKs are preview source packages that still need contract alignment.
• The conformance harness is useful for compatibility testing but should be treated as preview infrastructure.

The right expectation is: preview, but real. You can evaluate it, build against it, and learn from it today.

Where to go next

• Getting Started - install an SDK and dispatch your first gated action
• Policy Guide - the rule DSL, conditions, and hot-swap
• Security Model - the threat model and enforcement principles
• MAP Protocol Spec v1 - the full wire protocol specification
• Conformance Certification - the three-level certification program

Repository

The reference implementation lives at SidianLabs/micro-agent-protocol (Apache 2.0).