robinsaige

Check the dependency rating of an MCP tool before your agent calls it.

Humans build agents; agents call tools they've never seen. This is a thin, dependency-free client for Robin Saige — a neutral observatory that probes ~10,000 MCP servers and publishes what they actually are: alive or dead, honest or drifting, a real server or one of 1,300 costumes on a single operator. Ask it before you depend on a stranger's tool.

pip install robinsaige

Read a rating

import robinsaige

r = robinsaige.check("io.github.cyanheads/usaspending-mcp-server")
rating = r["rating"]
print(rating["cluster"])            # e.g. "mid-solo"
print(rating["distinctiveness"])    # 0 = typical, 1 = highly unusual
print(rating["reach"], rating["use"], rating["trust"])

Every figure is a percentile or class within the ~10,813-server census — a dated observation, not a warranty. What it does not claim is spelled out in rating["not_claimed"].

Find the right tool

Don't know the name? Search by what you need — ranked by rating, so trustworthy live servers come first.

res = robinsaige.find_tools("US tariff data")
for r in res["results"][:5]:
    print(r["cluster"], r["server"], r["matched_tools"])

Gate a tool call

Put a TrustGate in front of tool dispatch. It never silently allows a tool it couldn't rate — the default on a missing rating is warn, not allow.

from robinsaige import TrustGate, Blocked

gate = TrustGate()                        # sensible default policy

d = gate.assess("some.registry/name")
if d.ok:
    call_the_tool()                       # allow or warn
else:
    skip_it(d.reason)                     # blocked: dead-listing / unreachable / not-mcp / rate-limited

# or enforce — raises Blocked on a blocked cluster:
try:
    gate.guard("some.registry/name")
except Blocked as b:
    ...

Report how a call went (the tap)

Reading a rating and contributing one are the same install. Wrap a tool call in gate.tap(...) — it times the call and fire-and-forgets a report on exit. The observatory anchors every report to its own probe (a "worked" report on a server it saw dead is discarded) and reports do not move the verdict — they accumulate. Reporting failures are swallowed; the tap never breaks your agent.

with gate.tap("some/server", tool="lookup") as t:
    result = call_the_tool()
    t.ok = looks_good(result)     # optional; defaults to "no exception raised"

Policy is yours to widen or narrow:

gate = TrustGate(
    block={"dead-listing", "unreachable", "farm-costume"},   # refuse these
    warn={"stale-thin-solo", "auth-walled"},                 # warn on these
    on_missing="block",                                      # unrated = refused
)

Command line

robinsaige pulse                       # the whole ecosystem at a glance
robinsaige find tariff                 # discover a server by name fragment
robinsaige check io.github.you/mcp     # full rating (JSON)
robinsaige gate io.github.you/mcp      # ALLOW / WARN / BLOCK  (exit 2 if blocked)

What the clusters mean

dead-listing a registry entry pointing at a 404 · farm-costume many names, one operator · walled-opaque a login with no legible challenge · stale-thin old protocol, few tools · rich-fast-solo a large, quick, independent server. Browse them all at robinsaige.com/spectrum.

Honesty

Robin Saige is a signal, not a gate — it publishes its method and raw data so you can recompute and disagree. This client is stdlib-only on purpose: a trust tool should add zero dependencies. It is a dated observation, never a guarantee; auth-walled internals are unverified from outside.

MIT licensed.