About
How MCP Scoring works
We connect to your remote endpoint live, run 30+ protocol, tool and documentation checks, and hand you copy and paste fix prompts. Verified servers join the public MCP leaderboard.
What we test
The checks that decide whether an agent can use your server
Protocol compliance
Streamable HTTP and SSE negotiation, capability declarations, tools/list and resources/list behaviour.
Tool definitions
Names, titles, schemas and behavior hints checked against the MCP spec plus Claude, ChatGPT and Cursor limits.
Tool documentation
Whether every description says what a tool does, when to use it and what comes back, and every input is explained.
Server metadata
Name, title, description, version, website, icon and agent instructions from the initialize handshake.
Discoverability
An mcp.json or Server Card under /.well-known and a readable setup page when a person opens your endpoint.
Fix prompts
Every failing check becomes a precise prompt with spec links that a coding agent can apply in your repository.
How it works
From URL to A grade in three steps
- 01
Paste your endpoint
Any public Streamable HTTP or SSE MCP URL. Add an auth header if the server is private.
- 02
Get a scored report
Six sections, every check explained, with a letter grade and history across runs.
- 03
Apply the fix prompts
Paste them into your coding agent, deploy, and review again to climb the leaderboard.
Agents pick the tools they understand. Make yours the obvious choice.
For agents
Score MCP servers from inside your agent
MCP Scoring runs its own MCP server. Add it to Claude, ChatGPT, Cursor or VS Code to score endpoints, search the directory and pull fix prompts in conversation. It passes every check we run.
score_mcp_serversearch_mcp_directoryget_mcp_server_reportget_mcp_fix_promptlist_scoring_checks
Endpoint
https://mcpscoring.com/mcpClaude Code
claude mcp add --transport http mcp-scoring https://mcpscoring.com/mcpFAQ
MCP testing questions, answered
What is MCP Scoring?
MCP Scoring is a free tool that tests remote Model Context Protocol (MCP) servers and scores them from 0 to 100. It checks the handshake metadata, protocol behaviour, tool definitions and tool documentation, then ranks verified servers in a public MCP directory.
How do I test my MCP server?
Paste your server's Streamable HTTP or SSE endpoint, for example https://example.com/mcp, and press Score server. The scorer connects live, lists tools and resources, runs every check and saves a report in about 5 to 30 seconds.
What is a good MCP server score?
85 or above earns an A and means agents can discover, understand and call your tools reliably. Most servers lose points on missing behavior hints, undescribed inputs, absent output schemas and no server instructions.
Which servers appear in the MCP directory?
Only real MCP servers. A server is listed when its latest review completed an MCP handshake and exposed at least one tool. Websites, broken endpoints and servers behind failing auth are recorded privately but never ranked.
How do I raise my score?
Every failing check has a fix prompt with spec links. Copy it, or the Fix everything prompt, into Claude Code, Cursor or another coding agent working in your server's repository, deploy, then review again to track the change.
Are my auth headers stored?
No. Headers you add for a private server are used for that single review and are never written to the database or logs.
Can my AI agent use MCP Scoring?
Yes. MCP Scoring runs its own MCP server at /mcp. Add it to Claude, ChatGPT, Cursor or VS Code to score servers, search the directory and fetch fix prompts without leaving your agent.
Does it test local stdio servers?
The scorer tests remote servers over HTTP. To score a stdio server, expose it through a Streamable HTTP bridge on a public URL and review that endpoint.
Find out what agents
think of your server.
Free, no signup, results in under 30 seconds.
Score serverMCP Scoring never stores your auth headers.