SAM (Sovereign Agent Mesh) is an Apache‑2.0 licensed peer‑to‑peer overlay that lets autonomous AI agents share tools without exposing internal endpoints to the public internet. It addresses the fragmentation of agents running on clouds, on‑prem servers, laptops, Raspberry Pis and Android devices by providing a zero‑config, zero‑trust network built on the Model Context Protocol.
Nodes discover each other automatically, traverse NAT, and authorize every call cryptographically. A node is started with sam-node join and sam-node run; libp2p listens on UDP 5001 and TCP 5002, while the local MCP API defaults to port 8080. Authentication relies on an OIDC JWT that is converted into a Biscuit token whose claims become Datalog facts (e.g., user(...), group(...), client_peer_id(...)).
Authorization follows a strict default‑deny policy; access requires an explicit capability fact such as granted_service_exact(...). The request pipeline first checks ban/revocation caches, then executes two Biscuit authorizer passes—one for the node’s identity, one for the caller’s token—while binding the connection peer ID to prevent replay. The mesh is currently a beta testnet, not an official Google product.
- License: Apache‑2.0
- libp2p: UDP 5001, TCP 5002
- MCP API default: TCP 8080
- Auth: OIDC JWT → Biscuit token → Datalog facts
- Default‑deny, explicit capability facts required
- Two‑stage Biscuit pass + peer‑ID replay check
Why this matters
The source confirms that SAM provides a default‑deny, Biscuit‑based authorization layer over a libp2p‑backed P2P mesh, which means agents can enforce fine‑grained tool access without relying on external authentication services. This reduces the attack surface by keeping internal LLM endpoints and private APIs off the public internet. However, because the project is labeled a beta testnet and is not an officially supported Google product, any production deployment would need additional stability testing, integration work, and possibly further security audits before being trusted in critical workloads.