developer tools / agent safety
MCP Doctor
a CLI scanner for evaluating whether MCP servers are safe, clear, and agent-ready.
overview
MCP Doctor is an open-source CLI tool that scans Model Context Protocol servers and evaluates whether their tools are clear, safe, and ready for AI agents to use. It analyzes tool descriptions, input schemas, side-effect clarity, confusing tool pairs, dangerous unconstrained arguments, metadata-injection risks, and scoring signals that affect how reliably agents choose and call tools.
I built MCP Doctor as a practical linting layer for the growing MCP ecosystem, similar to how developers use ESLint or Lighthouse, but for agent tool quality. The project includes a deterministic scoring engine, detailed Markdown and JSON reports, benchmark scans across real MCP servers like GitHub MCP, Playwright MCP, Filesystem, Git, Memory, Fetch, and Sequential Thinking, plus documentation explaining scoring methodology, rule design, and safe real-server testing.
The key insight behind the project is that MCP servers do not just expose APIs. they expose an agent decision surface. tool names, descriptions, schemas, side effects, and overlapping capabilities all shape whether an AI agent behaves reliably.
why it matters
MCP Doctor is my strongest agent-safety and developer-tools project. it treats agent tools as an interface and safety surface, asking not only whether a tool works, but whether an AI agent can understand, choose, and call that tool reliably.
what i built
- built an open-source CLI scanner for evaluating MCP server tool quality.
- implemented deterministic scoring for descriptions, schemas, side effects, confusing tool pairs, risky arguments, and metadata-injection patterns.
- generated actionable Markdown and JSON reports with scores, findings, and top fixes.
- benchmarked real MCP servers including GitHub MCP, Playwright MCP, Filesystem, Git, Memory, Fetch, and Sequential Thinking.
- wrote scoring methodology, rule-design documentation, and safe real-server testing notes.
- framed MCP servers as agent decision surfaces, not just API wrappers.