A new proposal for the MCP servers repository introduces Signet, a middleware layer that cryptographically signs tool call requests and responses—enabling non-repudiable audit trails for AI agent actions.
As AI agents gain the ability to take real-world actions—sending emails, making purchases, modifying files—organizations need strong audit guarantees. Who authorized this action? When exactly did it happen? Has the record been tampered with?
Traditional logging can answer some of these questions, but logs can be modified. Signet proposes cryptographic signatures that create non-repudiable records: if an action was signed, you can prove it happened exactly as recorded, and you can prove who authorized it.
Signet operates as middleware that intercepts MCP tool calls at two points:
{
"tool_call": {
"name": "send_email",
"arguments": {
"to": "alice@example.com",
"subject": "Meeting tomorrow",
"body": "..."
}
},
"signet": {
"request_signature": "eyJ0eXAiOiJKV1QiLCJhbGciOiJFZDI1NTE5In0...",
"signer": "did:key:z6MkhaXgBZDvotDkL5257faiztiGiC2QtKLGpbnnEGta2doK",
"timestamp": "2026-03-30T11:37:25Z",
"nonce": "abc123"
}
}
The signature covers:
Signet supports multiple signature algorithms, with Ed25519 as the default for its performance characteristics:
signet:
algorithm: Ed25519
key_source: env:SIGNET_PRIVATE_KEY
require_signatures: true
allowed_signers:
- did:key:z6Mk... # Allowed agent identities
Key design decisions from the PR:
The PR outlines several scenarios where Signet adds value:
PR #3759 was opened on March 30, 2026 and is currently under review. The implementation includes:
Open questions from the PR discussion include key management recommendations, performance impact benchmarks, and whether response attestation should be mandatory or optional.
If merged, Signet would be one of the first community servers focused purely on security infrastructure rather than external service integration. It signals growing enterprise interest in MCP as a production-grade protocol, not just a developer convenience.
The DID-based identity model also aligns with emerging standards in the agent identity space, potentially enabling cross-protocol signature verification with systems like A2A and UCP.