Chainspeak

MCP server that lets LLMs read Ethereum chain data over JSON-RPC. Read-only, no keys, no signing.

Tools — capability clusters, not endpoint wrappers:

• chainspeak_get_chain_status — chain id + name, latest block, base fee, gas tiers (slow/standard/fast), "a transfer costs ~X ETH now", blob fee, sync flag, and the measured upstream batch cap
• chainspeak_get_account — balance, nonce, is_contract, EIP-7702 delegation, verified reverse ENS; takes an address OR an ENS name, resolved at the same block as the read
• chainspeak_get_token — ERC-20 metadata + total supply (no holder needed), optional holder balance, historical block support
• chainspeak_get_transaction — status, value, gas_limit + gas_used (+%), fee paid precomputed, confirmations, decoded method + ERC-20/721 transfers; failed txs get failure: {reason, method, method_note} via debug trace or honest eth_call replay; detail: summary|full|raw
• chainspeak_get_block — header, tx count, gas fullness, base fee; detail adds tx hashes or full transactions, paginated with steering truncation messages
• chainspeak_get_events — contract event logs by preset (transfers/approvals for an account or a whole token, raw by contract/topics); Transfer/Approval decoded, everything else honestly decoded: false; any range width (walked server-side in windows the provider accepts), paginated
• chainspeak_resolve_name — ENS both directions (name→address, address→name with forward verification), block-pinned, EIP-55 output

Conventions every tool follows: responses echo {chain_id, block_number}; addresses are EIP-55 checksummed; chain quantities are decimal strings with human-readable twins; blocks accept decimal, 0x-hex, or tags; errors are {category, retryable, message, hint} where the hint says what to change — deterministic provider rejections are never labeled retryable.

RPC note: historical-state queries need an archive-capable endpoint — the publicnode default is NOT archive; https://eth.drpc.org (free) is. drpc's free tier caps JSON-RPC batches at 3, which the server respects automatically.

The server does not probe for archive or trace support, and upstream has no archive or trace field. Both existed and both lied — each asked a capability question up front and read whatever error came back as the answer. On a free-tier quota message the archive probe failed closed (archive: false from an endpoint whose historical reads worked) and the trace probe failed open (claiming debug_traceTransaction support that was never demonstrated) — and the answers were cached, so they stayed wrong. Just make the call: pruned state comes back as HISTORICAL_STATE_UNAVAILABLE naming what to change, and get_transaction falls back to eth_call replay with the reason in failure.note.

0.3.0 — a breaking output change

failure.confidence is removed from chainspeak_get_transaction.

It was one word restating another. confidence could only ever agree with method, so it carried nothing a reader did not already have — and paying for it in every failed-transaction payload, in output a model has to read, is a bad trade for a label.

What it meant has moved into method, which now says how the reason was obtained AND how far it can be trusted:

replay-not-reproduced is new, and it is why this is a fold rather than a deletion: that case used to be method: "replay" with confidence: "none", and dropping the field without it would have made "no reason exists" read as "here is the reason, approximately". method_note is unchanged.

Requirements

• Node.js >= 20
• pnpm

Setup

pnpm install
cp .env.example .env   # set one RPC_URL_<CHAIN> per chain you want
pnpm build

Run over stdio (Claude Desktop, Claude Code, etc.)

Add to your MCP client config, e.g. claude_desktop_config.json:

{
  "mcpServers": {
    "chainspeak": {
      "command": "node",
      "args": ["/absolute/path/to/chainspeak-mcp/dist/stdio.mjs"],
      "env": {
        "RPC_URL_ETHEREUM": "https://eth.drpc.org",
        "RPC_URL_BASE": "https://base.drpc.org",
        "RPC_URL_ARBITRUM": "https://arbitrum.drpc.org"
      }
    }
  }
}

Add one RPC_URL_<CHAIN> per chain. Every tool then takes a chain parameter, and account, transaction and chain-status queries read all configured chains by default — so "what does this address hold?" answers across ethereum, base and arbitrum in one call. ENS names work on every chain but are always resolved on ethereum, so keep RPC_URL_ETHEREUM set unless you only ever pass 0x addresses.

Or with Claude Code:

claude mcp add chainspeak \
  -e RPC_URL_ETHEREUM=https://eth.drpc.org \
  -e RPC_URL_BASE=https://base.drpc.org \
  -e RPC_URL_ARBITRUM=https://arbitrum.drpc.org \
  -- node /absolute/path/to/chainspeak-mcp/dist/stdio.mjs

Run over HTTP

pnpm dev:http

Clients connect to http://localhost:3000/mcp; GET /healthz is a health probe.

Auth is optional: set MCP_AUTH_TOKEN (e.g. openssl rand -hex 24) and clients must send Authorization: Bearer <token>; leave it unset and the server runs open, logging a warning — do that only on a private network or behind reverse-proxy auth, since anyone who can reach the port can spend your RPC quota.

Or with Docker:

docker compose up --build

Configuration

Known chains: ethereum, base, op-mainnet, arbitrum-one, polygon, bnb-smart-chain, gnosis, linea-mainnet, sepolia, holesky. Short aliases (op, arbitrum, bnb, linea) and chain ids (RPC_URL_8453) also work. The server refuses to start if an endpoint reports a different chain id than the name it was configured under.

Development

pnpm dev     # stdio server from source
pnpm check   # lint + typecheck + dependency rules + tests

License

MIT