MCP Server Attestation
Ed25519-signed MCP tool manifests + spawn attestation. Layer-2 supply-chain hardening.
Ask AI about MCP Server Attestation
Powered by Claude Β· Grounded in docs
I know everything about MCP Server Attestation. Ask me about installation, configuration, usage, or troubleshooting.
0/500
Reviews
Documentation
Part of the StudioMeyer MCP Stack β Built in Mallorca π΄ Β· β if you use it
mcp-server-attestation
Layer-2 supply-chain hardening for Model Context Protocol servers. Ed25519-signed tool manifests, runtime spawn-attestation, default-deny argument sanitizer.
Direct response to:
- OX Security marketplace-poisoning, April 2026 β 9 of 11 MCP registries accepted malicious servers. Anthropic's published position: "expected behavior".
- CVE-2025-69256 β Serverless Framework MCP RCE via
child_process.exec()command injection. - CVE-2025-61591 β Cursor MCP RCE through OAuth-installed malicious server with spawn hijack.
This package provides what Anthropic chose not to: cryptographic verification of which tools a server is allowed to expose and which spawn calls it is allowed to make. It is a drop-in dependency, not a runtime replacement.
A note from us
We have been building tools and systems for ourselves for the past two years. The fact that this repo is small and has few stars is not because it is new. It is because we only just decided to share what we have built. It is not a fresh experiment, it is a long story with a recent commit.
We love building things and sharing them. We do not love social media tactics, growth hacks, or chasing stars and followers. So this repo is small. The code is real, it gets used, issues get answered. Judge for yourself.
If it helps you, sharing, testing, and feedback help us. If it could be better, an issue is more useful. If you build something with it, tell us at hello@studiomeyer.io. That genuinely makes our day.
From a small studio in Palma de Mallorca.
Packages (npm workspaces)
| Package | Purpose |
|---|---|
mcp-server-attestation (packages/lib) | Library: Ed25519 sign/verify, manifest schema, sanitizer, spawn attester, TOFU trust store. |
mcp-attest-cli (packages/cli) | CLI mcp-attest: keygen, sign, verify, inspect, fingerprint, check-pin. |
mcp-attest-demo (packages/demo-server) | Reference MCP server (stdio, spec 2025-06-18) exposing 5 tools that demonstrate the library. |
Install
npm install mcp-server-attestation
# CLI:
npm install -g mcp-attest-cli
# Reference MCP server:
npx mcp-attest-demo
Node 20+. No external crypto dependencies β uses node:crypto Ed25519 primitives.
Five-line server quickstart
import { verifyManifestStrict, attestSpawnStrict, type SignedManifest } from "mcp-server-attestation";
import signed from "./signed/manifest.json" assert { type: "json" };
// 1. At startup: prove the manifest you ship is the manifest you signed.
verifyManifestStrict(signed);
// 2. Before every child_process.spawn:
attestSpawnStrict(signed as SignedManifest, { command, args });
That is the entire integration. Two function calls, no SaaS, no daemon.
Tools (reference server mcp-attest-demo)
| # | Name | readOnlyHint | destructiveHint |
|---|---|---|---|
| 1 | attest_verify_manifest | true | false |
| 2 | attest_inspect_spawn | true | false |
| 3 | attest_generate_manifest_template | true | false |
| 4 | attest_sign_manifest | false | false |
| 5 | attest_keygen | false | false |
Annotations are honest: attest_sign_manifest and attest_keygen write files on disk so they are not read-only, but they do not destroy existing data so destructiveHint stays false. See docs/THREAT-MODEL.md for the per-tool capability table.
CLI
mcp-attest keygen --out-dir ./keys --name prod
mcp-attest sign --manifest manifest.json --private-key keys/prod.key --out signed.json
mcp-attest verify --signed signed.json --pin # TOFU pin
mcp-attest verify --signed signed.json --sigstore # opt-in Rekor cross-ref
mcp-attest inspect --signed signed.json --command /usr/bin/echo --arg "hello"
mcp-attest fingerprint --public-key keys/prod.pub
mcp-attest check-pin --server my-server --signed signed.json
verify exits with code 2 on bad signature, code 3 on pin mismatch.
MCP spec compatibility
| Spec version | Status |
|---|---|
| 2024-11-05 | parseable in manifest, not target of reference server |
| 2025-03-26 | parseable in manifest, not target of reference server |
| 2025-06-18 | full target |
The library is transport-agnostic. The reference server is stdio-only.
Security model
- Trust-on-First-Use is the default. The first time you verify a server, its public key is pinned to
~/.mcp-attest/trust.json(override:MCP_ATTEST_TRUST_FILE). Subsequent verifications reject any new key for the same server name withTRUST_PIN_MISMATCH. This catches the Cursor-style malicious-update vector. - No bundled trusted-keys list. This package does not act as a gatekeeper. If you want stronger assurance, opt into
--sigstoreto cross-reference the public-key fingerprint against the Sigstore Rekor transparency log. - Default-deny argument sanitizer.
shellSafeStringblocks every ASCII shell metacharacter, NUL, CR, LF, zero-width characters, BOM, RTL/LTR overrides, and Trojan-Source isolates. Allowlist behaviour requires the explicitregex/enum/prefix/literalrule kinds. - Canonical JSON is the signed surface. Re-serialisation cannot change the signed bytes.
What this package does NOT do (out of scope):
- Sandbox or containerise the server process.
- OAuth flow hardening (separate
mcp-oauth-shieldbuild). - Network egress control.
- Auto-patch existing servers.
Tests
npm install
npm run typecheck
npm test
Test corpus includes CVE-replay fixtures (packages/lib/tests/fixtures/cve-2025-69256-payloads.json, cve-2025-61591-payloads.json). The build is a regression check: every payload must be blocked.
Distribution
- npm publish via GitHub Actions OIDC with
--provenance. - Reference server submitted to mcp.so and FastMCP Directory.
- Library separately published as
mcp-server-attestation.
About StudioMeyer
StudioMeyer is an AI and design studio based in Palma de Mallorca, working with clients worldwide. We build custom websites and AI infrastructure for small and medium businesses. Production stack on Claude Agent SDK, MCP and n8n, with Sentry, Langfuse and LangGraph for observability and an in-house guard layer.
License
MIT, Copyright 2026 Matthias Meyer (StudioMeyer). See LICENSE.
