MEV-MCP

A local Model Context Protocol server that turns the entire MEV.fyi Research Hub into a hybrid-search, reranker-backed, citation-aware research base any Claude (or other MCP-compatible) agent can query on-demand.

---

What it does

Drop the MEV.fyi Papers CSV into the repo, run six Python scripts, build one Rust binary — and you have an MCP server that exposes 8 research tools over a fully offline, sub-50 ms hybrid (BM25 + semantic) index of every paper in MEV.fyi's curated collection.

Tool	What it does
search	Hybrid BM25 + cosine over BGE-small embeddings, fused with RRF. Optional cross-encoder rerank. MEV-alias query expansion.
get_paper	Full clean markdown + metadata for a paper, optionally section-filtered.
list_papers	Filter the catalogue by curated MEV topic or release date.
list_topics	The curated MEV vocabulary with paper counts.
cite	BibTeX entry for a paper.
summarize_paper	Structured TL;DR: title + abstract + key excerpts from intro/conclusion/discussion.
find_related	kNN over per-paper centroid embeddings — "more like this".
citations	Internal citation graph (cites / cited_by) built from arXiv-id and DOI regex matching across the corpus.

Why this stack

Hybrid Python + Rust, each used where it shines:

Stage	Language	Why
Download + Playwright	Python	playwright-python + playwright-stealth are the most mature toolchain for browser-driven scrapes.
PDF extraction	Python	pymupdf binds to MuPDF (C) — already C-fast.
Embeddings (one-shot)	Python	fastembed is ONNX; identical speed to Rust at batch.
MCP server	Rust	Single static binary, arrancs <50 ms, sub-ms queries with tantivy + sqlite-vec, no Python in the hot path.

Architecture

CSV -> 01_download -> 02_extract -> 02b_arxiv_enrich -> 03_chunk -> 04_embed
                                                                       |
                                            05_build_index <-----------+
                                                  |
                                                  +--> 06_enrich_topics
                                                  +--> 07_citation_graph
                                                  |
                                                  v
                                  corpus/index.sqlite + corpus/tantivy/
                                                  |
                                                  v
                                           mev-mcp (Rust)
                                                  |
                                                  v
                                        Claude Code / any MCP client

Install

Prerequisites: uv (Python), cargo (Rust ≥ 1.80), Chromium runtime libs.

git clone https://github.com/<your-user>/MEV-MCP.git
cd MEV-MCP

# Python deps (uv creates and locks the venv)
uv sync
uv run playwright install chromium

# Rust binary (~30 MB stripped release)
cargo build --manifest-path mcp-server/Cargo.toml --release

Build the corpus

Each step is idempotent — re-run after dropping a new paper to the CSV or to corpus/pdfs/ and only the new work happens.

uv run python ingest/01_download.py         # httpx + Playwright -> corpus/pdfs/
uv run python ingest/01b_retry_failed.py    # smart per-host retry for failures
uv run python ingest/01c_download_authed.py # optional, see docs/AUTH_COOKIES.md
uv run python ingest/02_extract.py          # pymupdf -> corpus/text/<id>.md + meta
uv run python ingest/02b_arxiv_enrich.py    # canonical abstracts/authors via arXiv API
uv run python ingest/03_clean_chunk.py      # semantic chunks -> corpus/chunks/<id>.jsonl
uv run python ingest/04_embed.py            # BGE-small embeddings (CPU, ~5 min)
uv run python ingest/05_build_index.py      # SQLite + sqlite-vec + Tantivy + paper centroids
uv run python ingest/06_enrich_topics.py    # curated MEV taxonomy on top of arXiv tags
uv run python ingest/07_citation_graph.py   # citation edges via arXiv-id / DOI regex

Expected output on a fresh run with the bundled MEV.fyi CSV:

Papers OK:       ~236 / 256  (gated SSRN/RG/Nature/ACM the rest)
Chunks:          ~8,500 (avg ~36/paper, p50 ~720 tokens)
Topics:          ~82 curated MEV labels
Citation edges:  ~340 internal paper↔paper
Corpus size:     ~330 MB total (mostly PDFs)
Binary:          ~30 MB stripped

Register the MCP

Project-scope (recommended)

Drop a .mcp.json at the repo root (paths must be absolute):

{
  "mcpServers": {
    "mev-mcp": {
      "command": "/absolute/path/to/MEV-MCP/mcp-server/target/release/mev-mcp",
      "args": ["--corpus", "/absolute/path/to/MEV-MCP/corpus"]
    }
  }
}

A template .mcp.json.example is included.

User-scope (all your projects can use it)

claude mcp add mev-mcp --scope user -- \
  /absolute/path/to/MEV-MCP/mcp-server/target/release/mev-mcp \
  --corpus /absolute/path/to/MEV-MCP/corpus

Verify:

claude mcp list
# mev-mcp: /path/to/mev-mcp --corpus /path/to/corpus - ✓ Connected

Use it

In any Claude Code session where the MCP is registered:

Show me what the corpus says about LVR mitigation strategies for AMM LPs.

The agent automatically picks search, optionally chains summarize_paper or find_related, and cites paper IDs. Use the tool names directly for explicit calls: mcp__mev-mcp__search, mcp__mev-mcp__find_related, etc.

Smoke test from the shell:

echo '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"search","arguments":{"query":"sandwich attack mitigation","k":3}}}' | \
  mcp-server/target/release/mev-mcp --corpus corpus

Auth for gated papers (SSRN, ResearchGate, …)

Some papers sit behind Cloudflare-protected logins. See docs/AUTH_COOKIES.md for the cookie-based recovery flow. For SSRN specifically Cloudflare blocks all automated paths even with valid cookies — the repo bundles a manual-download helper:

uv run python ingest/ssrn_manual_helper.py   # generates corpus/auth/ssrn-manual.html
# open the HTML in your browser, click each title, save PDFs to corpus/pdfs/manual/
uv run python ingest/ssrn_manual_import.py   # auto-matches by abstract_id / title

Stack

• Python 3.11+ ingestion: httpx, playwright, playwright-stealth, curl-cffi, pymupdf, tiktoken, fastembed, apsw, sqlite-vec, tantivy-py, orjson, slugify.
• Rust ≥ 1.80 server: tokio, rusqlite (bundled), sqlite-vec, tantivy, fastembed, clap, serde, anyhow.
• Models (auto-downloaded, ONNX, CPU): BAAI/bge-small-en-v1.5 (384d, ~30 MB) for embeddings; BAAI/bge-reranker-base (~280 MB) for optional cross-encoder rerank.

License

MIT.

The pipeline imports PyMuPDF, which is licensed under AGPL-3.0. When you run the ingestion scripts that import fitz, the combined work is subject to AGPL terms. If you need a non-AGPL combined work, swap the extractor for pypdfium2 (Apache-2.0) — it requires <30 minutes of refactoring in ingest/02_extract.py.

Source data: paper titles and URLs come from the MEV.fyi Research Hub. All PDFs are downloaded from their original publishers and remain subject to the rights of their respective copyright holders.