When an AI retrieves something from an archive, I want the connection to the source to remain inspectable. Which server supplied it? Has the response changed? Does it contain someone's original words, or an AI's later interpretation?
Verifiable MCP is a working pattern for attaching a digital signature and provenance to an MCP response. The signed material includes declarations about the source, rights and freshness, so those declarations travel with the content being checked.
A valid signature can establish that the response matches what the signing server issued. The reader still needs to evaluate the source and the claim. Signing a statement does not make it true.
Keep the source and its labels together
The distinction between an original contribution and an AI-written response matters. So do the rights attached to each. The pattern places those labels inside the signed material, where changing them can be detected.
For my own archive, a check can establish what my server supplied. For an index of other people's publications, it establishes what the index supplied. It does not prove that the original publisher wrote or approved the index's representation.
Tried with different kinds of public knowledge
The public repository documents reference implementations for #ensakidag, the SSWC book and the RISE/AI Sweden connector. They bring different situations: my own material, contributions by many authors, and an index of third-party publications.
That variation is useful. A common verification pattern needs to preserve the differences between sources, authorship and rights. Each server's current implementation and scope still need their own check.
Something you can inspect and test
The repository contains the specification, Python and JavaScript verifiers, a conformance checklist and published test vectors. The tests include altered and invalid responses that a verifier must reject.
Start with the explanation, run the published examples, then examine a reference server or try the pattern with a corpus you operate. The useful feedback is what the test reveals, including where the implementation falls short.
An experimental pattern with a defined scope
The signature covers the server response. It does not certify the answer an AI later writes from it. A client also needs to perform verification; simply receiving signed metadata is not the same as checking it.
The specification remains experimental. It describes its trust assumptions and limits, and makes conformance something to test rather than a badge to assume.