agent-evidence-mcp
An offline, read-only MCP server that decides whether a claim about an AI system holds against evidence bytes you supply, and replays the agent-evidence conformance vectors.
It's for agent builders, auditors and platform engineers who want their assistant to check evidence instead of trusting a report.
agent-evidence is the open record format; probityai checks and grades records against it, and this server runs those checks for an assistant.
Install the released package: pip install agent-evidence-mcp==0.2.6. It takes the packages of its own family (agent-evidence-verify, agent-evidence-vectors, agent-evidence-vocabulary, agent-evidence-admission, jcs-admit) as compatible-release ranges, so it installs beside their latest releases, and pins every third-party dependency exactly. The last one carries the jcs-admit and dsse Rust crates as wheels.
Quick start
Add it to Claude Code, then ask the assistant to run the sample claim:
claude mcp add agent-evidence -- uvx agent-evidence-mcp@0.2.6
Or install the Claude Code plugin: /plugin install agent-evidence --marketplace probityai/probity.
Or call a tool from Python, with no client installed beyond the package:
uvx --from agent-evidence-mcp==0.2.6 python - <<'PY'
import asyncio
from mcp.client import Client
from mcp.client.stdio import StdioServerParameters
async def main():
async with Client(StdioServerParameters(command="agent-evidence-mcp")) as client:
result = await client.call_tool("verify_sample", {"scenario": "contradicted"})
print(result.structured_content["decision"], result.structured_content["reason"])
asyncio.run(main())
PY
It prints:
contradicted selected_span_missing_from_record
A passage the report quotes is missing from the source capture, so the claim is contradicted. Use supported or not_established as the scenario for the other two decisions.
Nine tools, all read-only, offline and bounded: verify_claim, verify_sample, list_vector_corpora, run_vectors, explain_vector, lookup_term and check_artifact_digest, canonicalize_json and verify_dsse. Nothing is fetched, written or sent.
Status
Version 0.2.6. The tests run every tool through a real MCP client against the shipped corpus, and a mutation check confirms they fail when a tool's output is broken.
Documentation
| page | read it for |
|---|
| Tools | every tool, its arguments and what it returns |
| Install | Claude Code, Gemini CLI, Docker and plain stdio |
| Agent guide | a short index for agents using the server |
| Changing the repository | the commands to run and what never to hand-edit |
License
Apache-2.0.