protogrid
Sign inAdd to your agent
New: State of MCP servers, October 2026Listed in the Claude and ChatGPT directoriesClaude listingChatGPT listing

The MCP registry for agents

Let your agent choose its own MCP servers. On evidence, not luck.

protogrid measures 56,798 MCP servers so your agent does not have to guess: does it answer, does it need a key, how good is it, did it change. You set the rules once; your agent finds, judges and connects on its own.

send an emailquery a postgres databasebrowse a web pagesearch github issues

Free. No sign-up. No key. We never hold your credentials.

Configurations forClaude CodeCodexCursorVS CodeOpenCodeGemini CLIGooseany mcpServers JSON

Publish an MCP server? Get an email the moment it breaks, before your users find out. Free.

Claim your serversWhat owners get

The registry, measured · 2026-11-05

Not every server can be scored. We show you which ones.

38,901remote endpoints probed, every few hours
391,733tools indexed for search by intent
33,305servers scored on quality, every check public
0credentials held, tools called or proxies in the path

16,862 tool descriptions changed meaning since 2026-09-23: we record the before and after of every one. How scores work.

Read the October 2026 reportState of MCP servers, October 2026: protocol, OAuth, tools, dependencies and quality across the registry.

MCP servers are everywhere. Knowing which ones work is not.

A registry listing tells your agent a server exists. It does not tell it whether the endpoint answers, whether it needs a key, whether its tools changed meaning last week, or whether its dependencies carry known advisories.

Dead endpointsSign-in nobody mentionedTools that change meaningKnown vulnerabilitiesConfigs that never connect

Without protogrid

  • reads five READMEs to guess what each server does
  • copies a config from a blog post, then the wrong one
  • connects, gets a 401, asks you for a key it never named
  • cannot tell a maintained server from an abandoned one

With protogrid

  • one search by intent, ranked by relevance and quality
  • the connection class says if it needs a key or a sign-in
  • a configuration for your client, secrets as placeholders
  • trust and quality scores with every check shown

Autonomy, with judgment

Your agent picks its own tools. You set the rules once.

Autonomous agents rarely fail at reasoning. They fail at the boring part: picking a tool that actually works. protogrid turns every MCP server into evidence an agent can read, and every preference of yours into a rule it follows. No human picks, installs or babysits each tool.

01 Choose

It finds candidates by itself

The agent describes the job ("send an invoice") and gets servers ranked by relevance and quality, across every tool in the registry.

02 Judge

It reads the evidence, not the marketing

Live probes, scores and history tell it which server answers, which needs a key, which changed last week. It drops what fails your rules.

03 Connect

It plugs in and gets to work

A ready configuration for its client; servers without auth need nothing, keys stay in your vault, OAuth asks you once.

Your rules, in one call
search("send an invoice",
  class=["R0", "R1"],  # reachable alone
  min_quality=75,      # good or strong
  exclude_quality_flags=["known-vulns"],
  owner_verified=True) # a real owner

Put it in your agent's instructions, the SDK or a plain API call. Every server it picks from then on meets your bar, and you can see why.

Answers right nowNeeds a key or a sign-inQuality 0-100Trust 0-100Tools changedKnown advisoriesOwner verified2026-07-28 protocol

Browse the registry

MCP servers for the jobs agents do most

EmailPostgreSQLGitHubSlackNotionJiraStripeWeb searchBrowser automationWeatherCalendarGoogle MapsAgent memoryKubernetesFilesystemAll use cases

Humans use the web. Agents use the API.

Add protogrid in one line

Our own MCP server is remote, needs no key and speaks the 2026-07-28 protocol. Listed in the Claude directory as a connector and as a plugin with two skills, and in the ChatGPT plugins directory.

REST API: GET /v1/search, /v1/servers/{name}TypeScript SDK (MIT)Python SDK (MIT)
claude mcp add --transport http -s user protogrid https://api.protogrid.dev/mcp

Then ask: "use protogrid to find a server that can send an email".

[mcp_servers.protogrid]
url = "https://api.protogrid.dev/mcp"

~/.codex/config.toml

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "protogrid": {
      "type": "remote",
      "url": "https://api.protogrid.dev/mcp",
      "enabled": true
    }
  }
}

~/.config/opencode/opencode.json

{
  "mcpServers": {
    "protogrid": {
      "url": "https://api.protogrid.dev/mcp"
    }
  }
}

~/.cursor/mcp.json

{
  "servers": {
    "protogrid": {
      "type": "http",
      "url": "https://api.protogrid.dev/mcp"
    }
  }
}

.vscode/mcp.json

{
  "mcpServers": {
    "protogrid": {
      "type": "http",
      "url": "https://api.protogrid.dev/mcp"
    }
  }
}

mcpServers JSON

For MCP server owners

Know when your MCP server breaks. Before your users do.

Agents already find your server on protogrid. Claim it and we watch it for you: hourly probes, an email when a tool changes or a check fails, and a public "owner verified" mark that agents can filter on.

Claim your servers, freeEverything owners get

Sign in with GitHub. One digest a day at most, one-click unsubscribe. Owner verified is never an audit and never a ranking boost.

  1. Step 1 · today

    Check it

    One probe of any URL: quality checks and what the Claude and OpenAI directories require. No account.

    Check a server
  2. Step 2 · every deploy

    Gate it in CI

    The GitHub Action checks your server after each deploy and fails the build on a regression.

    Get the Action
  3. Step 3 · all the time

    Claim and monitor it

    Owner verified, hourly probes, alerts and webhooks for 3 servers. Free.

    Claim your servers

Questions

Everything else is in the docs, and every page has a Markdown twin for agents.

I publish an MCP server. What do I get?

Claim it with GitHub or DNS and get owner verified, hourly probes, email alerts, signed webhooks and the README badge, free with 3 servers monitored hourly. Claim your servers.

Can my agent choose servers without asking me?

Yes, within the rules you set: connection class, minimum quality, no known advisories, a verified owner. Servers that need a key use the ones you already gave it; OAuth asks you once.

Is protogrid free?

Yes. Search, descriptors, configurations, the MCP server and checks need no account. A free account adds an API key with higher limits.

Do you see my keys or tokens?

Never. protogrid is a registry, not a gateway: secrets stay placeholders and your agent connects directly.

How are scores calculated?

From signals anyone can observe, with published weights. No audit is implied; every check shows its evidence.

How is this different from the official MCP registry?

We mirror it every hour, then add what an agent needs to choose and connect: live probes, search by intent, scores and configurations.

The whole MCP ecosystem, for the agents that use it and the people who build it.

Building an agentLet it choose and connect servers on its own, within your rules.Add protogrid to your agent
Publishing a serverGet told when it breaks, and show agents it is yours.Claim your servers, free