For operators

Get your MCP server measured.

Vouch runs real invocation trials against MCP servers and publishes what it finds. If you operate one, here’s exactly what that means for you — what gets probed, how to opt out, what verification actually changes, and how to get discovered.

What Vouch does to your server, exactly

Read-only, by default. Mutating tools are skipped and marked untested, never guessed at from a self-reported readOnlyHint — Vouch doesn’t trust that signal for anything, since it’s exactly the kind of self-report the whole project exists to distrust. For the deep-probe corpus, side-effect classification is done by hand, with the reasoning recorded and published alongside the tool.

Bounded concurrency. Never more than one in-flight request against your server at a time, and a hard ceiling of 4 concurrent calls across the whole run regardless — higher only against a self-hosted instance Vouch itself operates, or with your explicit consent, and always labelled as such when it happens.

Announced. Every probe identifies itself with a real, identifying User-Agent string. A small number of unannounced spot-checks exist purely to detect divergence between announced and unannounced behaviour — they never feed a published score, and they stay inside the same rate limits and read-only rules as everything else.

Full mechanism detail — the exact opt-out channels, what’s honoured today versus proposed — is at /probing-standard.

How to opt out

Three ways, checked in this order before every probe:

A /.well-known/mcp-probing.json manifest on your server’s origin with {"allow": false}.
A robots.txt disallow — User-agent: * (or any agent whose name contains vouch) with Disallow: /.
Email — the address is in every probe’s User-Agent string. A manual deny-list entry takes effect before the next run.

This isn’t a hypothetical: 22 servers had already opted out via robots.txt alone by the time of the first registry-wide census (see /research). Opting out doesn’t hide your server or penalize it — see the next section.

The measurement dial

Every server sits at exactly one of three states, and moving between them is entirely up to you:

You do	What Vouch shows
Opt out	Operator declined probing — a visible, counted state, not a blank and not a penalty.
Do nothing	A shallow-tier score — schema integrity, and a sample call where one was attempted.
Verify ownership	Eligible for the full deep-probe battery — see the next section for exactly what that means.

Declined renders with the same visual weight as not yet tested — no warning colour, no flag styling. Absence is stated, never implied.

What verification concretely gets you — and no further

Verification does	Verification does not
Make your tools eligible for the full behavioural battery instead of shallow-tier signals	Improve your score. A verified server with bad behaviour still scores badly.
Let you supply test credentials or a sandbox endpoint — the only route to measurement for gated (Class C) servers	Move you up any ranking. Deep-tier results rank against deep-tier results; a shallow-tier tool is never pushed down for being shallow.
Produce a report card with component-level detail and per-trial evidence, instead of a one-shot observation	Cost anything, ever. Free, permanently — not a promotional rate.

Stated as narrowly as that on purpose: verification buys measurement, not position. Full detail on how ranking within tiers works is in the methodology’s ranking rule, and its prohibitions are what this table is restating in plain language. Verifying your own server is a form on its report card — search for it at /search.

How do I get my tools discovered?

The honest answer isn’t “list on Vouch” — Vouch doesn’t host, sell, or list servers. It’s this:

Publish to the official registry (registry.modelcontextprotocol.io) — neutral, jointly backed, and the source Vouch itself syncs from. Expose /.well-known/mcp/server.json on your own domain so it can be auto-discovered there too.
List in more than one registry. No single one covers the ecosystem: Smithery (~7,000+ servers), Glama (tens of thousands, auto-indexed from GitHub), mcp.so (~20,000+), MCP.Directory (auto-pulls from GitHub within 24h, claimable), PulseMCP. Different audiences, different mechanisms — being in one and not the others is a common, avoidable gap.
Get measured. Verify (previous section), so whoever’s evaluating your server sees a number produced by someone with no commercial stake in it, not a description you wrote about yourself. A measured score is also what a downstream project’s CI check reads when it decides whether to depend on your server.
Put the result where people already look. Every report card has a ready-to-paste badge at the bottom — copy the markdown straight from your tool’s or server’s page. The badge renders at the same weight whatever the score says (no red, no warning colour), links back to the full report card, and is available whether or not the score is good — it is never a verification perk:
```
[![Vouch score](https://vouch.tools/api/tools/{id}/badge.svg)](https://vouch.tools/tools/{id})
```
A server-level variant (tool count and median score, honest about not cherry-picking your best tool) is on every server page at /api/servers/[id]/badge.svg.

The aim here is to be genuinely useful, including about other services — a page that answers a real question honestly is worth citing; a page that just says “list with us” isn’t.

What Vouch will never do

Charge for a score, a badge, verification, or an expedited review.
Run a certification programme. A number and its evidence, not a seal.
Let a private scan produce a private result that quietly enters the public index, in either direction — private scanning stays private.

Back to the search, the full methodology, or the probing standard.