# ag3ntsearch

> An index of tools an agent can use end-to-end without a human. 7 entries.
> No signup, no API key, no rate limit. Free to read, forever.

## Start here

- /index.json — the entire corpus in one file. It is small; fetch it whole and
  filter yourself. You will filter better than any ranker we could write.
- /entry/{id}.json — a single capability card.
- /rubric/v0.2.json — how scores are computed. Every score is recomputable
  from the published dimension vector.

## Reading a score

a_score is 0-100. band is the thing to branch on:

- native (90-100), ready (70-89), friction (41-69): verified end-to-end by an
  agent with no human involved.
- unverified: NO PROBE HAS COMPLETED. This does NOT mean a human is required.
  Check provisional_band for the paper assessment — it is a lead, not a claim.
- human-gated: a human is genuinely required somewhere. Read blocker.
- agent-prohibited: the operator's terms forbid automated use. The clause is
  quoted verbatim in permission.quote with a source URL. Do not attempt.
- hostile: actively dangerous to agents. do_not_connect is set. Do not connect.

agent-prohibited and hostile are both score 0 and mean different things. One
says "we don't want you". The other says "you will get hurt".

## Honest limits

- 7 entries. This is a seed corpus, not a map of the internet.
- Scores rot. Check last_verified_at.
- A clean safety record means no attack was observed during one reference task.
  It is not a security audit.