research-mcp

A local Model Context Protocol server that makes a personal paper library searchable from any MCP-capable client (Claude Desktop, Cursor, VS Code).

Ask "what do my baseline papers say about positional encoding?" and get answers grounded in the PDFs on your own disk, with page-level citations.

The problem

Research context is scattered. Papers live in one folder, experiment runs in MLflow, notes in a third place. Questions that span those sources — "which of my runs used the configuration from the paper I read last month?" — can't be answered without manual digging.

This server exposes each source as a tool. The language model decides at runtime which one to query and how to combine the results.

Architecture

Two programs, one SQLite file. They never talk to each other.

PDFs ──► ingest.py ──► library.db ◄── server.py ◄──► Claude Desktop
         (offline)                    (stdio MCP)

ingest.py extracts text, splits it into overlapping chunks, computes embeddings and writes everything to SQLite. You run it when you add papers.

server.py exposes read-only tools over stdio. Claude Desktop starts it automatically. It contains no LLM — retrieval is deterministic Python, synthesis happens in the host.

Tools

Setup

Requires Python 3.12+ and uv.

git clone https://github.com/<user>/research-mcp
cd research-mcp
uv sync

Organise papers by project. The folder structure becomes queryable metadata:

data/papers/
├── thesis/
│   ├── baselines/
│   └── related-work/
└── general/

Index them:

uv run ingest.py            # index new or changed files
uv run ingest.py --force    # re-index everything
uv run ingest.py --stats    # show current contents
uv run ingest.py --no-arxiv # skip metadata lookup (offline)

Register the server in claude_desktop_config.json:

{
  "mcpServers": {
    "research": {
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/research-mcp", "run", "server.py"]
    }
  }
}

Verify without a model in the loop:

npx @modelcontextprotocol/inspector uv run server.py

Design decisions

Ingestion is separate from the server. The server is read-only and loads the embedding model lazily, so Claude Desktop starts in milliseconds instead of waiting for PyTorch. A write path inside the server would also mean an embedding model resident in memory for every session.

No print() anywhere in the server. With stdio transport the MCP protocol occupies stdout — a single stray print corrupts the message stream. All logging goes to stderr.

References are stripped before chunking. Bibliographies are dense clusters of domain vocabulary with no propositional content; leaving them in hijacks semantic search.

Filtered vector search overfetches. The KNN query is unaware of the metadata columns and returns the k globally nearest chunks; the project filter is applied afterwards. Without overfetching (k = limit * 8) a filtered query can return almost nothing. This is the standard pre- vs post-filtering tradeoff in ANN search.

Titles come from font size, not from the first line. Paper title pages frequently carry licence notices above the title. Taking the largest text span on page one is substantially more reliable. When an arXiv ID is present, the arXiv API overrides the heuristic entirely.

Content hashing drives re-indexing. Each PDF is fingerprinted with SHA-256, so ingest.py is idempotent — unchanged files are skipped, changed files are replaced along with their orphaned vectors (virtual tables are not covered by ON DELETE CASCADE).

Stack

Python MCP SDK (FastMCP) · SQLite + sqlite-vec · sentence-transformers (all-MiniLM-L6-v2) · PyMuPDF · httpx

Status

Working: PDF ingestion, semantic search, project/area scoping, arXiv metadata lookup.

Planned:

• Test suite covering chunking and retrieval, running in CI

• Hybrid retrieval (BM25 via FTS5 + dense, combined with reciprocal rank fusion)

• MLflow integration replacing the get_run_metrics placeholder

• Note enrichment: a tool that gathers evidence for rough notes, keeping retrieved facts and model-generated synthesis in separate, clearly labelled sections

Licence

MIT