Provenance for the agents you are about to adopt
AI Registry
Every agent carries a passport recording what its build actually discloses (model, framework, tools, data sources, hosting, residency and permission scope) and who said so.
The passport
One record per agent, listing what the source states about the build and marking who stated it. A value is Verified when the marketplace assessed it on a certification page, Disclosed when the publisher asserted it, and Unknown when nothing was said. Those three are never collapsed into one.
- Creator, primary model, agent framework, tools and MCP servers
- Data sources, integrations, hosting model and data residency
- Permission scope, compliance certifications and deployment readiness
- The listing version, the source it came from, and when it was captured
Evidence risk is not a security rating
The risk band measures how much of the build you cannot see before you deploy, not how dangerous the agent is. It starts at the attestation level and moves on how much of the build the source can disclose but does not. An agent with a thin record scores badly even if it is perfectly safe, because the point is what you can check.
Comparison and hand-off
Filter by function, source, deployment, evidence tier, provenance, pricing and evidence risk; put candidates side by side; and share the exact result set as a link. Facet counts exclude their own selection, so the numbers beside a filter tell you what would happen if you applied it.