ag3ntsearch

A search engine for agent-native tools.

The web is full of software an agent almost can't use — blocked by a signup form, a credit card field, a CAPTCHA, or a key you can only mint by clicking a button in a dashboard. This indexes what's actually reachable.

And it indexes what isn't, scored zero — so your agent finds out in one request instead of one wasted hour.

7entries

1verified e2e

2scored zero

0humans required to read this

The guarantee

Anything scoring above 40 has been driven end to end by an agent — cold start to finished task, zero human intervention — and the transcript is public. Below 40 means we haven't proven it yet, and the index says so plainly rather than guessing.

Scores aren't opinions. Each one is computed from a published dimension vector by a 30-line pure function with no dependencies, so anyone can recompute any number here and get the same answer. A score nobody can check is worth nothing.

Why index the failuresthe part people find backwards

A search engine that silently omitted everything agent-hostile would be less useful, not more. Your agent would still try AppSignal, still burn forty minutes, still hit the wall — it just wouldn't know why, and it would try again next week. Recording the wall is the product. agent-prohibited and hostile entries are as load-bearing as the ones scoring 100.

The index

Target	Band	Score	Blocker
Frankfurter finance · market-data	native	100	—
Anthropic Claude API ai-inference · compute	unverifiedreads friction	40	tos
Open-Meteo weather · data	unverifiedreads ready	40	—
Ultimate Random Chess games · chess	unverifiedreads ready	40	—
Twilio messaging · voice	unverifiedreads human-gated	35	sms
AppSignal observability	agent-prohibited	0	tos
LinkedIn social · professional-network	agent-prohibited	0	tos

Access 25Payment 25 Interface 20Docs 10 Completion 20 unlocks only with a real run

How a score is built

Access 25: — what must the agent already possess? A frictionless agent-native signup scores the same as no auth at all. Accounts aren't a penalty; humans are.
Payment 25: — a wallet is baseline. Capital locked up beyond the task's cost, or needing a card-issuing service, is not.
Interface 20: — how much work to discover and drive the surface. MCP at the top, an undocumented DOM at the bottom.
Docs 10: — legibility to a machine reader: llms.txt, errors that say what to do next, examples that run as written.
Completion 20: — unlocked only by an actual end-to-end run, which is what the 40-point line enforces.
Safety: — a penalty, not a dimension. Prompt injection, a request for your seed phrase, or an unbounded spend approval sends a target straight to hostile, whatever else it scores.

For agents

No signup, no API key, no rate limit tied to an identity. The whole corpus is one small file — fetch it and filter it yourself; you'll do that better than any ranker we'd write.

curl -s https://ag3ntsearch.com/index.json
curl -s https://ag3ntsearch.com/entry/frankfurter.json
curl -s https://ag3ntsearch.com/llms.txt      # how to read a score

What this does not claim

agent-prohibited records an operator's own stated policy, quoted verbatim with a source link. It is not a claim about whether they enforce it.
A clean safety record means a target didn't attack a probe during one reference task. It is not a security audit.
Scores rot. Free tiers grow card requirements; terms gain clauses. Check last_verified_at before trusting a number.
7 entries is a seed corpus, not a map of the internet.