MCP server that exposes workspace tools to AI agents for browsing, creating, and updating markdown documents organized into sections
AIDotMd provides 9 tools with consistent naming (verb_noun pattern) and reasonable descriptions. All tools have input schemas with type definitions and parameter descriptions. However, output schemas are not formally documented in the code, parameter descriptions lack depth (missing constraints, ranges, format specs), and error handling is minimal. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles. Descriptions average ~100 chars, below the 194-char baseline for A+ tools. Missing guidance on idempotency, retry safety, and recovery paths.
Finalize a streaming write and save the accumulated content to the database. Must be called after stream_write to persist the content. Until this is called the content only exists in memory and will be lost on server restart.
Create a new markdown document in a section.
Create a new section. Can be top-level or nested under a parent section.
Get a document's full markdown content.
List documents in the workspace.
List all sections in the workspace, including nested ones.
Output schemas not documented. Tools return dicts but LLMs cannot see field types, required fields, or structure. Forces LLMs to infer output shape from examples.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools have clear risk profiles (READ_ONLY vs WRITE) but LLMs cannot see them. Agents may retry destructive ops or treat reads as safe.
Parameter descriptions lack constraints. 'title' has no length limits; 'limit' in search_docs has no min/max; 'content' has no size cap. LLMs will pass unbounded values.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
Full-text search across all documents in the workspace.
Write a chunk of content to a document in real-time. The chunk is broadcast immediately to any browser tabs viewing the document so the user can watch the content appear live. Call commit_stream when done to save the final content to the database.
Update an existing document.
No error recovery guidance. Tools return dicts with 'error' field but do not guide LLM on next steps (e.g., 'Document not found. Try list_documents() to find valid IDs.').
stream_write + commit_stream require two-step workflow but lack idempotency guarantees. If commit_stream is called twice, does it duplicate content? Unclear to LLMs.