Capability Resolution MCP server
Commit Nov 12, 2026Enables AI agents and gateways to discover which servers can perform specific tasks through signed, expiring recommendations.
Also known as XLSV777/capability-resolution
- Stars
- 0
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 12, 2026
- License
- MIT
What is Capability Resolution MCP?
Capability Resolution enables AI agents and gateways to discover which servers can perform specific tasks through signed, expiring recommendations.
Summary generated from the project's README
About Capability Resolution MCP server
Helping AI agents and gateways find the right tools for specific tasks is the primary function of capability-resolution. This open specification allows a client to ask which servers can perform a particular job and receive a list of up to ten candidates. The system operates over HTTP, where clients discover resolver endpoints and public keys through a well-known URI.
To ensure security, each recommendation is signed using JWS with EdDSA over Ed25519 and expires within 48 hours. Developers can use the provided Node.js packages, such as the capability-resolver-client, to implement these checks. A unique feature in version 0.3 is the signed tool fingerprint, which allows a client to detect if a server's tools have changed since the resolver last checked them. This mechanism provides advice rather than direct instructions, leaving the final decision to connect to the client or user.
What you can do with it
- 1Discovering specialized tool serversAsk a resolver to find servers capable of sending transactional emails without requiring a sign-in.
- 2Verifying tool integrityCheck if a server's available tools have changed since the last time a resolver inspected them.
Written from the project's README. Every use case is checked against a quote from it.
README
View on GitHubCapability Resolution
An open specification for one question an AI agent, client or gateway can ask over HTTP: "which servers can do this job, and did they answer when last checked?"
A resolver answers with a short list of candidates. Each candidate is signed by the resolver (JWS, EdDSA over Ed25519) and expires within 48 hours. The client verifies every answer, fails open when anything is wrong, and always keeps the final decision: a candidate is advice, never an instruction to connect.
- Spec: spec/capability-resolution-0.3.md, also at https://7it.co.il/standard/capability-resolution/
- Version: 0.3 (open specification, versioned), 2026-11-05. New in 0.3: a signed tool fingerprint (
tfp), so a client can detect that a server's tools changed since the resolver looked. Previous: 0.2. - Editors: 7IT Solutions. We welcome co-editors from other resolver, gateway or client projects.
- License: the spec text is CC BY 4.0 (LICENSE-SPEC.md); the code is MIT (LICENSE).
- Status: an independent proposal, for use with the Model Context Protocol and other tool protocols. Not part of the Model Context Protocol specification, and not endorsed by Anthropic, the MCP maintainers, the Linux Foundation or any vendor.
- Trademarks: Model Context Protocol and MCP are trademarks of their respective owners.
How it works
- Discover.
GET https://<resolver>/.well-known/capability-resolverreturns the resolver's endpoints and public keys. - Ask.
GET <resolve>?need=send%20a%20transactional%20email&constraints=no_sign_inreturns at most 10 candidates. - Verify. The client checks each token's signature, issuer and lifetime, and drops anything that does not verify.
- Decide. The client or its user decides whether to connect. The resolver only recommends.
- Compare (0.3). When a token carries
tfpand the client connects, it lists the server's tools and compares their fingerprint with the signed one before the first tool call. A mismatch means the tools changed since the resolver looked.
Server owners can publish a short declaration file at /.well-known/capabilities.txt (section 13): what their servers
do, whether they need sign-in or payment, a contact, and which resolvers may list them. A resolver must respect a deny
and still check each server itself: the file is a claim, not proof. See examples/capabilities.txt.
Any party may run a resolver. How a resolver chooses and orders candidates is its own method and is outside the spec.
What is in this repository
| Folder | What |
|---|---|
| spec/ | The specification text. |
| packages/resolver-js | capability-resolver-client: a Node 18+ client with no dependencies (conformance client-1), and the capability-resolution-check CLI that tests a resolver (resolver-1) or a declaration file (server-1). |
| packages/test-vectors | Signed test tokens, valid and invalid, with expected verdicts and the public test key. |
| packages/adapter-ai-sdk | Example: resolveTool() for the Vercel AI SDK. |
| packages/adapter-langchain | Example: CapabilityResolverTool for LangChain JS. |
| examples/ | A declaration file and an llms.txt summary of the spec for agents. |
| proof/ | The exact text of each version with its RFC 3161 and OpenTimestamps proofs. |
npm package names: capability-resolver-client, capability-resolver-ai-sdk, capability-resolver-langchain and
capability-resolution-test-vectors (see PUBLISHING.md). To work from source, clone this repository:
cd packages/resolver-js
npm test
node bin/capability-resolution-check.js https://<resolver origin>
node bin/capability-resolution-check.js --server https://<your host>
Contributing
Issues and pull requests are welcome: questions, corrections, a second implementation, or a request to co-edit. See CONTRIBUTING.md.
Capability Resolution MCP questions
What is the purpose of Capability Resolution?
It is an open specification that allows an AI agent, client, or gateway to ask which servers can perform a specific job and when they were last checked.
How does a client verify the server recommendations?
The client verifies the signature, issuer, and lifetime of each token, dropping any that do not verify.
Can a resolver force a client to connect to a specific server?
No, the resolver only provides recommendations; the final decision to connect rests with the client or its user.
How can a client detect if a server's tools have changed since the last resolution check?
By using the signed tool fingerprint (tfp) introduced in version 0.3, the client can compare the server's current tools against the signed fingerprint.