Atinamos Evidence MCP — pre-purchase verification evidence for autonomous AI buyers #848

MotionFil-AI · 2026-08-26T15:59:51Z

MotionFil-AI
Aug 26, 2026

Pre-submission Checklist

This post follows the MCP community guidelines

What would you like to share?

We’ve built Atinamos Evidence MCP, a public read-only MCP intended to sit inside an autonomous agent’s procurement flow before it spends money on a machine service.

The idea is simple: an AI buyer discovers a paid service, queries Atinamos for published execution evidence about that service, then evaluates that evidence against its own procurement or spending policy.

The MCP currently exposes four tools:

lookup_evidence(endpoint) — retrieve published verification evidence for a service route
service_history(endpoint) — retrieve timestamped published observations
search_services(query, limit) — search services represented in the published evidence corpus
evaluate_policy(endpoint, policy) — evaluate the available evidence against buyer-supplied rules

A buyer policy can include requirements such as a minimum number of successful fulfilments, maximum failed fulfilments, paid evidence being required, maximum evidence age, and maximum observed price.

The result of policy evaluation is eligible, not_eligible, or insufficient_evidence. This is deliberately not a universal Atinamos trust or safety verdict.

Our core principle is:

Atinamos provides the evidence. The buyer owns the policy and decision.

### Important scope boundary

MCP v1 currently reads only the selected, sanitised Atinamos evidence that has been approved for public publication. It does not query or expose our complete private machine-service registry, operational database, unpublished observations, or internal commercial intelligence.

This is an intentional production boundary while we prove the public interface, evidence semantics and buyer-policy behaviour.

Therefore, if the MCP returns no evidence for a service, that means only that no published Atinamos evidence is currently available through MCP v1 for that route. It does not mean that the service does not exist, is unsafe, or has never appeared in Atinamos’ private research.

The server is published in the official MCP Registry as:

uk.co.atinamos/evidence

Version 1.0.0

We’d particularly welcome feedback from people building autonomous purchasing/procurement agents: is a pre-spend evidence check like this useful in the decision loop, and what other evidence or policy controls would you expect?