neuronto

Find the MCP servers, skills, agents and APIs that can actually do a task.

One call searches this index and every other public Agentic Resource Discovery (ARD) registry at once, fusing the rankings, so you get one answer across the federation instead of a single catalogue's view.

No account, no API key, no signup. Python 3.9+. No dependencies.

Command line

pip install neuronto

neuronto "read a PDF and extract tables"
neuronto tools "convert currency"     # individual tools, not whole servers
neuronto stats                        # what the index holds, measured
neuronto dead                         # endpoints that stopped answering

Options: --limit N, --kind mcp|api|skill|agent, --local (do not federate), --json, --key KEY.

From code

from neuronto import find_resource, find_tool, registry_stats, liveness

out = find_resource("post a message to Slack", limit=5)
for r in out["results"]:
    print(r["displayName"], r["url"], r["score"])

# Tool-level search. Names come from each server's own tools/list.
for t in find_tool("convert currency", limit=5):
    print(t["tool"], "on", t["server"])

find_resource(query, *, limit, kind, federate, api_key)

Returns {"results": [...], "federation": {...} | None}. federation["registries"] says which registries answered and how long each took. A search sent with api_key is not recorded at all.

find_tool(query, *, limit)

Results carry tool (the server's own name for it), server, endpoint, score, verified.

registry_stats()

The measured state of the ecosystem: what share of probed endpoints answer, how many expose the tools they claim, median response. The window the numbers come from and their limitations travel in the payload, so a reader who quotes a number also gets its caveats.

liveness(*, dead=False, since=0, limit=500, cursor=0)

Liveness observations, including the endpoints that stopped answering. Free to use, redistribute and build on. No key, no attribution required. If you run a registry, dead=True is the useful half.

publish(*, endpoint="", domain="")

Get an endpoint or a domain indexed. Verified rather than trusted: the endpoint has to complete a handshake, or the domain has to serve a manifest that parses.

About the score

score is semantic relevance only. It is not a trust, safety or quality rating and must not be presented as one. verified is separate and reports what was observed by fetching.

Publishing your own resources

This package is the read side. To build and validate an ARD manifest for your own domain, use ard-publish.

Privacy

There is no column for an IP address, and a keyed search is not recorded. Detail, written from the running code: https://neuronto.com/privacy.

Pointing it elsewhere

Set NEURONTO_BASE to any other ARD registry's base URL; they implement the same search interface. The list is at https://neuronto.com/ard-registries.

Licence

Apache-2.0.