Why

Specs for blockchain protocols live across eleven different upstream repos on three different forges. Every time you need to look up EIP-4844, BIP-340, CIP-25, or which fork shipped PUSH0, you're tab-hunting through GitHub. This MCP indexes them all locally with FTS5 ranking — 1,767 proposals across 10 chains plus addresses for 19 protocols on Ethereum, Arbitrum, Base, Optimism, Polygon, and more — so your agent answers with the actual spec text, not a hallucinated paraphrase.

Install

Requires: Python 3.11+ · uv (provides uvx) · ~500 MB free disk for the index · git on PATH (used by --sync).

Step 1 — build the index (one-time, ~2 min, ~500 MB in ~/.cache/web3-docs-mcp/)

uvx web3-docs-mcp --sync

Step 2 — register the server with your agent

The launch command is identical across clients:

uvx web3-docs-mcp

claude mcp add web3-docs -- uvx web3-docs-mcp

{
  "mcpServers": {
    "web3-docs": {
      "command": "uvx",
      "args": ["web3-docs-mcp"]
    }
  }
}

codex mcp add web3-docs -- uvx web3-docs-mcp

pip install web3-docs-mcp
# or:
pipx install web3-docs-mcp

git clone https://github.com/dioptx/web3-docs.git && cd web3-docs
uv sync
uv run python server.py --sync   # build index
uv run python server.py          # run stdio server

Restart your agent, then try "Use web3-docs to look up EIP-1559."

What you can ask

Tools

Sources

11 upstream repos, all synced via --sync:

Fork mappings come from ethereum/execution-specs plus canonical Bitcoin soft-fork activations (P2SH, SegWit, Taproot, …).

Contract registry covers: aave, across, chainlink, compound, create2_deployer, curve, ens, erc4337, gnosis_safe, lido, maker, multicall, oneinch, permit2, seaport, uniswap, usdc, usdt, weth.

Why not …

…just gh search or WebFetch each spec on demand? You'd burn tokens on HTML markup and pay a network round-trip per query. web3-docs indexes everything once into local SQLite + FTS5 — sub-millisecond ranked search, plain-text bodies, no rate limits, works offline.

…one MCP per chain? You'd manage eleven separate servers and your agent wouldn't know which to call. One unified tool with a single resolve_proposal entry point lets the model find the right doc by concept (e.g. "blob transactions" → eip-4844) rather than guessing the source.

…ask the model directly without an MCP? Models hallucinate spec details — wrong fork, wrong gas costs, wrong opcode numbers. This server returns the actual upstream text with metadata (status, fork, activation date) so the agent can quote it verbatim.

…use a vector DB? Spec corpora are small (≈ 1.7K docs), domain vocabulary is precise (PUSH0, BLOBHASH, taproot), and exact-term matching beats embeddings here. FTS5 gives BM25 ranking with zero infrastructure.

Configuration

Troubleshooting

"Index is empty" on any tool call. You haven't run --sync yet. Run:

uvx web3-docs-mcp --sync

uvx: command not found. Install uv: curl -LsSf https://astral.sh/uv/install.sh | sh.

Want to free disk space? Source repos (~/.cache/web3-docs-mcp/repos/) can be deleted after sync; only proposals.db is needed at runtime. Re-run --sync to update.

Stale data? Re-run --sync — it does a fast git pull and reindexes incrementally.

Development

git clone https://github.com/dioptx/web3-docs.git && cd web3-docs
uv sync --extra test
uv run pytest                  # 98 tests, BDD + unit
uv build                       # build wheel + sdist

Status

v0.2.0 — adds Cardano CIPs, Tezos TZIPs, Sui SIPs (10 chains, 1,767 proposals). SQLite + FTS5, FastMCP stdio transport. See CHANGELOG.md for release history.

License

MIT — see LICENSE.