Taskflow Doc Search

doc-search MCP server

A small MCP server written with FastMCP. It gives an AI coding agent structured access to this workspace's documentation and demo tickets — the local, public stand-in for the Jira/Confluence side of the original production workspace. Five tools are annotated read-only (embeddings maintains a derived local cache); the sixth can only write ticket *drafts* (see below).

Tools

Draft-first writes

propose_ticket mirrors the production workspace's rule for outward writes (there: Jira changes start as drafts a human approves). The boundary is enforced by the tool, not by politeness: it writes only new files under tickets/drafts/, always with status: draft, it refuses to overwrite an existing draft, and it is annotated as a write, so clients ask the user before calling it. tickets/ is the pre-TF-115 archive; list_tickets never shows drafts as tickets. A human reviews and promotes a draft with scripts/promote-draft.ps1 (or .sh), which files an Issue instead of moving the file into tickets/. Set AI_WORKSPACE_PROJECT to the workspace Project number, or pass -Project N / --project N, to also add the Issue to the board. Without a Project, the script prints the Issue URL and notes that it was not added to a board. The draft is preserved. The doc-search server stays credential-free and never calls GitHub.

Three rungs of document search

1. search_docs counts exact term hits with title/heading/body weights. It is predictable for known terms, but repetition can dominate ranking.
2. search_docs_semantic uses TF-IDF cosine similarity over whole documents. It down-weights common terms and balances multi-term queries. Both of these first two rungs are bag-of-words models: neither understands synonyms. They use only the standard library and rebuild their vectors per call.
3. search_docs_embeddings uses sentence-transformers over ATX heading sections, embedding the document title, heading trail and body. It can retrieve paraphrases and synonyms, but relevance is model-dependent. Cosine ranking uses pure Python. A content-hash fingerprint invalidates the derived .index/ cache when documents or the model change.

• Local documents: searches never modify the source documents. Embeddings writes a derived cache; propose_ticket writes drafts under the rules above.
• No credentials: keyword and TF-IDF need no network. Embeddings downloads its model once, then runs fully locally.
• Client-agnostic: plain MCP over stdio.

Setup

cd mcp\doc-search

python -m venv .venv .venv\Scripts\python -m pip install -r requirements.txt pytest

Optional embeddings dependencies (includes torch):

.venv\Scripts\python -m pip install -r requirements-embeddings.txt

DOC_SEARCH_EMBEDDING_MODEL overrides the default sentence-transformers/all-MiniLM-L6-v2. DOC_SEARCH_EMBEDDER=fake selects only a deterministic hashed-token test double, without semantic understanding.

Evaluate the synced corpus with python scripts/run_eval.py; add --write to record the dated results in EVAL.md. The eval is not run in CI.

Run the tests

.venv\Scripts\python -m pytest

The tests build a throwaway fixture workspace, so they pass whether or not sync has populated projects/.

Register it

Codex — add to ~/.codex/config.toml:

[mcp_servers.doc-search]

command = 'C:\path\to\ai-workspace-kit\mcp\doc-search\.venv\Scripts\python.exe' args = ['C:\path\to\ai-workspace-kit\mcp\doc-search\server.py']

Claude Code:

claude mcp add doc-search -- C:\path\to\ai-workspace-kit\mcp\doc-search\.venv\Scripts\python.exe C:\path\to\ai-workspace-kit\mcp\doc-search\server.py

The server finds the workspace root from its own location ( /mcp/doc-search/server.py); set AI_WORKSPACE_ROOT to override.

Install the distribution

Once published, install with pip install taskflow-doc-search (or pip install "taskflow-doc-search[embeddings]" for sentence embeddings). For a local checkout, pip install . from this directory works today. Run taskflow-doc-search with AI_WORKSPACE_ROOT set to the absolute path of the workspace to search. It serves MCP over stdio; no command arguments are needed. The existing python server.py launcher and its automatic workspace-root detection remain supported.

Release instructions: PUBLISHING.md. The included server.json is prepared for an optional MCP registry listing and has not been published.