mcprated

Agent-readable index of
every MCP server.

Built for LLMs to discover, vet, and choose tools at runtime — not for humans to browse. Open ruleset, daily-updated, deterministic. Live MCP endpoint + static JSON catalog. Eight agent-shaped tools.

# Claude Code / Claude Desktop — one line
claude mcp add --transport http mcprated https://mcp.mcprated.workers.dev

# Cursor / Cline / Continue — paste into config
{"mcpServers": {"mcprated": {"url": "https://mcp.mcprated.workers.dev"}}}

servers rated

scoring 100

median composite

v1.3.2

rule set

What we measure

Four axes. Twenty signals. One composite score, 0–100. Each axis answers a single question a developer asks before installing an MCP server.

Reliability

Will it work and keep working?

CI present, no floating SDK deps, recently maintained, has releases, semver tags, release notes.

Documentation

Can a stranger figure this out?

Substantive README, install instructions, tools documented, code examples, external docs link.

Trust

Safe to depend on?

Commercial-friendly license, security policy, properly tagged for discovery.

Community

Are people caring for it?

Multiple contributors, responsive issues, recent merged PRs, history beyond initial commit, contributing guide.

Top 10 today

#	Repo	Score	Rel	Doc	Trs	Com
1	awslabs/mcp	86	70	100	73	100	Python · ★8,951
2	microsoft/playwright-mcp	84	100	100	36	100	TypeScript · ★32,014
3	modelcontextprotocol/python-sdk	84	100	100	36	100	Python · ★22,878
4	modelcontextprotocol/typescript-sdk	82	90	100	36	100	TypeScript · ★12,341
5	GLips/Figma-Context-MCP	80	100	100	18	100	TypeScript · ★14,614
6	oraios/serena	80	100	100	18	100	Python · ★23,828
7	eyaltoledano/claude-task-master	77	90	100	18	100	JavaScript · ★26,988
8	sooperset/mcp-atlassian	77	100	80	27	100	Python · ★5,096
9	github/github-mcp-server	74	70	100	27	100	Go · ★29,507
10	googleapis/mcp-toolbox	74	70	100	27	100	Go · ★14,936

Full catalog: index.json · Per-server: servers/<owner>__<repo>.json

Manifesto

001 / Open by default

The full ruleset is YAML in a public repo.

No "trust us" scoring. Anyone can read every signal definition, fork the linter, dispute a result with evidence, propose a change via pull request. Trust comes from auditability, not authority.

002 / Deterministic over clever

Twenty binary signals. Same input, same score. Forever.

No machine-learned weights. No LLM-as-judge in the core score. No popularity contests. A signal either passes or it doesn't, and the rule that decided is published. Reproducibility is the whole point.

003 / Versioned, never silently changed

Every score carries the rule-set version that produced it.

When the bar rises, we tag a new major version, publish a delta report explaining what shifted and why, and keep old versions queryable for at least a year. Embedded badges don't break under your feet.

004 / The data is the product

Daily snapshots in GitHub Releases under CC-BY-4.0.

Pull the JSON, build your own catalog, train your own classifier, embed badges. Attribution is the only ask. The site is the demo; the data is the artifact.

005 / Maintainers deserve specifics

Every failed signal links to the exact fix.

"Score 60 because we don't like you" is a non-answer. We tell you which signal failed, why it failed for your repo specifically, and how to make it pass. Disagreement gets a public reply, never silence.

For agents

MCPRated is built agent-first. The site is a UI; the API is JSON files served from GitHub's CDN. No auth, stable URLs, daily refresh.

# Discovery — start here. Lists every endpoint and what it contains.
GET https://mcprated.github.io/mcprated/llms.txt

# Catalog index — every server with composite, axes, kind, capabilities, distribution
GET https://mcprated.github.io/mcprated/index.json

# Per-server full lint detail (kind, subkind, capabilities, all signal results)
GET https://mcprated.github.io/mcprated/servers/<owner>__<repo>.json

# Transparency: what we filtered out and why
GET https://mcprated.github.io/mcprated/excluded.json

# Daily snapshot archive (full history)
GET https://github.com/mcprated/mcprated/releases

Each server JSON carries kind (server / client / framework / tool / ambiguous), subkind (integration / aggregator / prompt-tool / agent-product), capabilities[] from a versioned taxonomy, and distribution. Use these to answer "what does this server do" and "is it actually a server" without re-reading the README.

Coming next: agent-shaped endpoints (/api/v1/find?capability=…, /vet/<slug>, /by-capability/<cap>.json) and @mcprated/mcp-server as an installable npm package — query the catalog directly from your agent.

Embed your badge

If you maintain an MCP server, drop this into your README. Always reflects the latest score.

# Markdown
[![MCPRated](https://mcprated.github.io/mcprated/badges/<owner>__<repo>.svg)](https://mcprated.github.io/mcprated/)

# Example: microsoft/playwright-mcp
[![MCPRated](https://mcprated.github.io/mcprated/badges/microsoft__playwright-mcp.svg)](https://mcprated.github.io/mcprated/)

Renders as:

Replace <owner>__<repo> with your repo path using double underscore as separator (e.g. microsoft__playwright-mcp). The badge auto-updates daily — no URL pinning to maintain.

Status

v0.2 — agent-first core. Linter + classifier + capability taxonomy live, daily snapshots running, 32 servers indexed. Coming next: agent-shaped endpoints (find, vet, by-capability) and @mcprated/mcp-server npm package.

GitHub → Methodology → Changelog → Ruleset (YAML) → Snapshot archive →