Audit an MCP server's tool interface for agent-safety problems before you connect an agent to it.
mcp-probe is an audit tool, not a production MCP server. It exposes 22 tools across multiple example servers (clean_notes_server.py, unsafe_notes_server.py, fake_server.py, etc.). The clean_notes_server example demonstrates best practices with tool annotations and confirmation patterns, but the unsafe examples intentionally violate safety patterns. Most tools lack complete parameter descriptions, and many are duplicated across files. The tool set is heterogeneous (notes, jobs, shell commands, exports) without clear composition. Descriptions vary from detailed (clean_notes) to minimal (generic 'A tool that does something useful'). No output schemas are documented. This is a reference/testing tool, not a production server.
Create a note at a path that does not exist yet. Fails if something is already there, so it never overwrites.
Delete one job from the queue. The job cannot be restarted afterwards.
Delete one job from the queue. The job cannot be restarted afterwards.
Delete a note from the store. The note cannot be recovered afterwards.
Delete a note. The note is removed from the store and from the history, and cannot be recovered.
Export.
A tool that does something useful.
Return the current status of the queue and the number of jobs in it.
Duplicate tool definitions across files (list_notes, read_note, delete_note, get_status appear 2-3 times). This violates single-responsibility and confuses tool discovery.
Minimal descriptions in test fixtures (banner_server, chatty_server, paginated_server): 'A tool that does something useful' is generic and provides no LLM guidance on when/why to call the tool.
Unsafe examples (write_note, run_shell, export_all) lack confirmation patterns and have minimal descriptions. While intentional for audit purposes, they demonstrate anti-patterns without clear warnings in tool metadata.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | <=2025-11-25 | v2 |
Return the current status of the queue and the number of jobs in it.
List every note in the store, newest first.
List every note in the store, newest first.
A tool that does something useful.
Return the full text of a single note.
Return the full text of a single note.
A tool that does something useful.
Run a shell command in the notes directory.
Full text search over every note. Returns note ids and a matching snippet for each hit.
Full text search over every note. Returns note ids and a matching snippet for each hit.
A tool that does something useful.
Send the daily digest email to every subscriber.
Replace the body of a note that already exists. The previous body is kept in the store's history for thirty days.
Write a note.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields without knowing response structure.
Tool names lack verb clarity in some cases (read_thing, list_roots, first_tool, second_tool). Generic names like 'first_tool' and 'second_tool' provide zero semantic information.