MCP Server Review
Review a Model Context Protocol server for handshake, capabilities, schema quality, and operator-facing gaps. Use when shipping or consuming an MCP server, reviewing mcp.json, or preparing a conformance scan.
Copy the file into your agent runtime. Cursor loads personal skills from ~/.cursor/skills/mcp-server-review/SKILL.md and project skills from .cursor/skills/mcp-server-review/SKILL.md. Claude Code looks in .claude/skills/mcp-server-review/SKILL.md.
MCP Server Review
When to use
An MCP server is about to ship, be consumed, or be listed. Triggers: review this MCP server, check our tools/list, is this spec-compliant.
What it does
Walks handshake, capability advertisement, and tool/resource/prompt schemas the way an operator would, and produces prioritized fixes. Complements a live scan; this skill works from source, docs, or a captured initialize/tools/list transcript.
Steps
- Confirm transport (Streamable HTTP or stdio) and the initialize → notifications/initialized sequence.
- List advertised capabilities. Flag capabilities advertised but unimplemented, and implemented but unadvertised.
- For each tool: name, description, input schema, destructive vs read, auth assumptions.
- Flag missing
requiredfields, untyped objects, and descriptions that invite prompt injection (run whatever the user said). - Check error handling: unknown method, malformed JSON, missing session id.
- Prioritize fixes: handshake breakage first, then schema, then hygiene.
Output
Grade guess: A–F (label as static review, not a live scan) Blockers: - ... Should-fix: - ... Nice-to-have: - ...
Guardrails
Do not claim a live handshake passed unless a real initialize exchange is in the input. If only source is available, label findings Observed (in code) vs Hypothesized. Never recommend shipping a server that executes unsandboxed shell from tool arguments.