AP
AgentPacts
Independent lab
Join the waitlist
Lab online

Popularity is not quality.

AgentPacts is the independent lab that actually runs AI tools, then tells every agent which one to trust. The PageRank and AdWords for the agent economy.

7 servers scored | 1 lab run | traces public
The problem

AI agents pick tools the way we pick restaurants from ads: confidently, and without evidence. Discovery ranks marketing copy. Nobody runs the tools to check. We do.

They

They rank what agents choose.

We

We test whether the choice was right.

AgentPacts in Claude Code

Which tool fits this task?

Illustrative Claude Code session: a user asks AgentPacts which tool to choose for a noisy street interview with accents and local-only processing. The response checks those specific constraints rather than relying on a general transcription score, and does not recommend an unverified match.
Illustrative terminal session, not a captured live run.
How the lab works
01
Fixed tasks with known answers

Every task checks against a ground truth, verified by string match, not a model.

02
The same test for every server

Identical fixtures, identical conditions. Only the server changes.

03
Every run leaves a public trace

Timestamped tool-call logs you can read yourself, for every score.

04
Scores are never for sale

Money can reorder ties. It can never change a number.

Don't trust us. Reproduce it.

The marketplace

PageRank plus AdWords for the agent economy.

Vendors can buy prominence, but only among statistically tied tools that clear the quality bar. The bid is the offer; the audit trail is the invoice.

We once refused a bid from a failing server, and we published the refusal.

Coming soon

today: transcription and todos, 7 servers scored.

next: translation, web extraction, and every category where agents choose.

categories open as they're lab-verified, never before.

Join the waitlist

Be first in line when your category opens.