ainet Connect your agent
ainet-probe · hourly crawl
initialize → tools/list · no tool calls

The agent web, measured. Not just listed.

ainet probes every endpoint in the official MCP registry, every hour, and records what actually answers: status, latency, auth, tools. Your agent reads the ranking through one MCP tool, find(task).

Connect your agent See live data
Why ainet exists

Registries list every agent endpoint that exists. A listing can't tell you which ones answer, how fast, or behind what login. A probe can.

Existence
Listed in a catalog
Probed every hour
Uptime
Claimed by the publisher
Measured, probe after probe
Access
Found out when the call fails
Open, OAuth or paid, known upfront
Ranking
A flat list of names
Scored 0 to 100 on evidence
The probe· one crawl, three passes

From listed to ranked.

01 · Discover

Pull every endpoint from the registry.

Each crawl starts from the official MCP registry. On every host we also look for an A2A agent card.

02 · Probe

Shake hands. List tools. Time it.

An MCP initialize, then tools/list. No tool is ever called and no credentials are sent.

03 · Score

Rank by evidence.

Reach, speed, verify and safety add up to a score from 0 to 100. Every result carries the breakdown.

Endpoints—
Answered—
Scored 80+—
Live index· what your agent gets back

One question. Ranked answers.

This is the index your agent queries. Type a task and read each result's status, latency, score and matched tools. Pick one to vet it.

Try

Injection signals

static heuristics over tool definitions · signals, not verdicts

—endpoints flagged
override
Wording that tells the model to ignore its instructions, or that poses as a system message.
hidden
Invisible characters or buried comments that a reviewer would not see.
exfil
Wording that asks the model to pass secrets, credential files or its prompt into a tool call.
When a definition matches
Recorded
tool name, kind and an excerpt
find(task)
the endpoint is left out
check(url)
still visible, flags included
What the scan found

Search by task, in plain words.

Matched by meaning and by keyword. Results come back ranked, each with its status, latency, score and the tools that matched.

  • Search by meaning, in any language
  • Down or flagged endpoints never returned
  • At most two results per host
Per endpoint· what gets measured

Measured, not claimed.

Every endpoint gets the same probe and the same record. This is everything we keep, and everything your agent can read back.

— score 80+
Evidence score0 to 100, the sum of reach, speed, verify and safety.
— answered
ReachabilityWhether the MCP handshake got an answer in the latest crawl.
UptimeThe share of all probes that got an answer, over time.
— median
Handshake latencyTime to the first handshake response.
— endpoints
Auth requiredA 401 or 403 is recorded. Nothing more is attempted.
— endpoints
OAuthWhether the 401 advertises OAuth metadata.
— endpoints
Payment requiredHTTP 402 responses, recorded as paid.
— tools
Tool listTool names and the full count, as the server returns them.
— changed
Definition driftWhether tool definitions changed since we first saw them.
— flagged
Injection signalsStatic kinds: override, hidden, exfil. Signals, not verdicts.
— hosts
A2A agent cardWhether the host also serves one.
Probe historyThe latest 48 probes per endpoint, with status and latency.
Connect

One endpoint. Two tools.

No signup, no API key, no SDK. Point your agent at the MCP endpoint and it can start asking.

Step 1

Add the endpoint

One command, or the same URL in any MCP client config.

claude mcp add --transport http ainet /mcp
Step 2

Ask for a task

Your agent describes the job. The ranked endpoints come back with their evidence.

Tool call
find({
  task: "send an email",
  limit: 5,
  open_only: true
})
Step 3

Vet before you allow

One URL in, its full record out: status, auth, uptime, tools and flags.

Tool call
check({
  url: "https://host.example/mcp"
})
Questions

Straight answers.

Do I need an account or an API key?

No. There is no signup, no API key and no SDK. Add the MCP endpoint and your agent can call find and check.

Does the probe ever call my tools?

No. It sends an MCP initialize, then tools/list, and stops. It never calls a tool and never sends credentials. If your endpoint answers 401 or 403, it is recorded as auth-required and nothing more is attempted.

Can I opt my server out?

Yes. The probe identifies itself with the user agent ainet-probe. Block that user agent to opt out.

Is a flag a security verdict?

No. Flags are static heuristic signals found in tool definitions: override, hidden and exfil. They show you where to look. They are not verdicts. Flagged endpoints are left out of find results and stay visible through check.

Is the search semantic?

Yes. Describe the task in plain words, in English or Spanish: it is matched by meaning with embeddings and by keyword, and the endpoints come back ranked by evidence.

Do you crawl more than one registry?

No, not yet. Today the source is the official MCP registry, and on each host we also check for an A2A agent card. ARD catalogs and x402 marketplaces are next. They are not available yet.

Give your agent the measured web.

One MCP endpoint, two tools. Every endpoint in the official MCP registry, probed each hour and ranked by what answers.

Connect your agent See live data
No signup, no API keyNever calls a toolProbed again every hour