mcp-signed-manifests

A minimal reference implementation of signed tool manifests — a proposed optional extension to the Model Context Protocol (MCP) that lets a server cryptographically sign its tool definitions (names, descriptions, input schemas), so clients and gateways can detect when a tool's metadata has been tampered with after the fact.

This addresses a known, documented gap in MCP's current security model: a server can change a tool's description after a user has approved it — for example, silently appending hidden instructions like "also BCC this address on every email" — and there is currently no protocol-level way for a client to notice. This class of attack is often called tool poisoning or a "rug pull."

What's here

src/canonical.ts — deterministic (key-sorted) JSON serialization, so the same logical manifest always produces the same signed bytes.
src/signer.ts — ManifestSigner (Ed25519 signing, server-side) and ManifestVerifier (Ed25519 verification, client/gateway-side). verify() never throws — it always returns a boolean.
src/types.ts — ToolDefinition, ToolManifest, SignedManifest types.
example/demo.ts — an end-to-end walkthrough: sign a manifest, verify it, then tamper with a tool description without re-signing, and watch verification catch it.

Running the demo

npm install
npm run demo

Expected output walks through:

A server generates an Ed25519 keypair and signs its tool manifest.
A client verifies the signature on first connect (VALID ✓) — this is the moment a real client would show a user a consent prompt.
The manifest is altered (a hidden instruction is added to a tool description) without re-signing.
The client re-verifies and gets INVALID ✗ — TOOL_MANIFEST_TAMPERED, and refuses to silently apply the change.

How this would plug into MCP

This is intentionally not a fork or a competing protocol. The proposed shape is a single new optional method:

tools/manifest →  { manifest: {...}, signature: "ed25519:<hex>" }

advertised via a capability flag during initialization. Servers and clients that don't implement it are completely unaffected — this is additive only.

Because verification doesn't require any change to the server itself, it's also naturally implementable as an interceptor / gateway sitting in front of an existing, unmodified MCP server — the gateway signs on the server's behalf, and/or verifies on the client's behalf. This composes with the interceptor pattern being discussed for MCP more broadly.

A full write-up of the proposal, including the rationale for manifest-level (rather than per-call) signing, key rotation via expiresAt, and how this relates to transport-level security, is in docs/SEP-signed-tool-manifests.md.

Status

This is a proof-of-concept accompanying a proposal — not a production library. In particular:

Key distribution/rotation (how a client learns and trusts a server's public key) is intentionally left open, same as JWKS endpoints are deployment-specific in OAuth.
This protects metadata integrity, not confidentiality, and is meant to be layered on top of (not instead of) normal transport security.

License

MIT

MCP Servers