arb-dex-mcp

Live cross-DEX crypto prices for your AI agent — per-venue pool price, pool liquidity

and the gross cross-venue spread on 6 EVM chains, read straight from on-chain pool state.

Chains: BSC · Polygon · Arbitrum · Base · Avalanche · Optimism.

Venues: PancakeSwap (v2 + v3), Uniswap v3, SushiSwap, QuickSwap, Biswap, ApeSwap, BaseSwap,

Trader Joe, Pangolin — every v2-style pool and every v3 fee tier separately, because a

$12k 1% pool and a $19M 0.01% pool are not the same quote.

Nothing is modelled, estimated or backfilled. Every payload states its own block number and

carries its own scope note, so an agent that quotes a figure also has the caveats attached to it.

Works with no API key against a free hourly public snapshot. Docs · npm

Quickstart

Nothing to clone or build. Your MCP client fetches the package. Requires Node 18+.

Claude Desktop

`claude_desktop_config.json` — macOS `~/Library/Application Support/Claude/`,

Windows `%APPDATA%\Claude\`:

{
  "mcpServers": {
    "arb-dex": {
      "command": "npx",
      "args": ["-y", "arb-dex-mcp"],
      "env": {
        "RAPIDAPI_KEY": "your-rapidapi-key-here"
      }
    }
  }
}

Restart Claude Desktop; the six tools appear under the connectors icon.

Drop the `env` block entirely to run keyless — the server still starts and the

free-snapshot tools still answer.

Claude Code

claude mcp add arb-dex --env RAPIDAPI_KEY=your-rapidapi-key-here -- npx -y arb-dex-mcp

Keyless:

claude mcp add arb-dex -- npx -y arb-dex-mcp

Then `/mcp` to confirm it connected.

Cursor

`~/.cursor/mcp.json` (global) or `.cursor/mcp.json` (per project):

{
  "mcpServers": {
    "arb-dex": {
      "command": "npx",
      "args": ["-y", "arb-dex-mcp"],
      "env": {
        "RAPIDAPI_KEY": "your-rapidapi-key-here"
      }
    }
  }
}

Cursor → Settings → MCP shows the server and its tools once the file is saved.

Any other MCP client

Same three facts: command `npx`, args `["-y", "arb-dex-mcp"]`, transport stdio,

optional env `RAPIDAPI_KEY`.

Try it

Real prompts, and the real shape that comes back. Payloads below were measured live on

2026-08-15; they are trimmed for width but nothing is invented.

1. "What is WBNB/USDT trading at on every BSC venue right now?"

`get_prices` reads every pool holding the pair — v2 pairs and each v3 fee tier separately —

at one stated block:

{
  "pair": "WBNB/USDT",
  "network": "bsc",
  "chainId": 56,
  "blockNumber": 116151268,
  "pricesByVenue": [
    { "venue": "pancake",        "surface": "v2", "feeBps": 25,  "price": 611.7347, "tvlUsd": 56834814.36 },
    { "venue": "biswap",         "surface": "v2", "feeBps": 10,  "price": 610.8278, "tvlUsd": 415733.94 },
    { "venue": "apeswap",        "surface": "v2", "feeBps": 20,  "price": 611.1244, "tvlUsd": 3659.06 },
    { "venue": "pancakeV3:1",    "surface": "v3", "feeBps": 1,   "price": 610.5684, "tvlUsd": 18596810.32 },
    { "venue": "pancakeV3:5",    "surface": "v3", "feeBps": 5,   "price": 610.6531, "tvlUsd": 4941099.44 },
    { "venue": "pancakeV3:25",   "surface": "v3", "feeBps": 25,  "price": 610.6304, "tvlUsd": 52534.35 },
    { "venue": "pancakeV3:100",  "surface": "v3", "feeBps": 100, "price": 609.1025, "tvlUsd": 12206.21 }
  ],
  "bestBuy":  { "venue": "pancakeV3:100", "price": 609.1025 },
  "bestSell": { "venue": "pancake",       "price": 611.7347 },
  "midSpreadBps": 43.22,
  "crossDex": {
    "grossSpreadBps": 0,
    "grossUsd": 0,
    "optimalInput": { "amount": 0, "token": "WBNB", "usd": 0 },
    "buyVenue": "-",
    "sellVenue": "-"
  },
  "liquidity": { "venues": 7, "totalTvlUsd": 80856857.67 },
  "source": "rpc"
}

Read the two spread numbers against each other. The raw mid spread is 43 bps — and the

gross capturable spread is 0. The 609.10 quote lives in a $12k pool; the size that would

actually clear it moves the price past the gap before you get there. A tool that reported only

the 43 bps would be handing an agent a number it cannot trade. This one reports both, and

`optimalInput` is where the honesty lands.

2. "Show me the cross-DEX spreads on Base — are any actually capturable?"

`get_spreads` sweeps a whole chain and ranks by gross USD at the optimal size, not by

headline basis points:

{
  "network": "base",
  "chainId": 8453,
  "scannedPairs": 11,
  "opportunities": [],
  "found": 0,
  "filters": { "minSpreadBps": 10, "minVenueTvlUsd": 1000, "minGrossUsd": 0.01, "limit": 5 },
  "ranking": "gross USD at the optimal trade size, NOT raw spread — a large spread with a tiny optimal size is not an opportunity",
  "scope": "GROSS cross-venue spread from live pool state, BEFORE gas, MEV and any slippage beyond the optimal size. Not a profit estimate and not trade advice. Venues below the liquidity floor are excluded because a spread against a dust pool is an artefact, not an opportunity.",
  "elapsedMs": 2847
}

`found: 0` is a real answer and it is the common one. Eleven pairs scanned, nothing cleared

the floor. Venues under $1,000 TVL are dropped outright. When rows *do* come back, each carries

`capturable`, `warning` and `shallowestSideTvlUsd` so a big basis-point number cannot mislead on

its own. This tool will tell your agent there is nothing there — which is the whole point of

asking it.

3. "How much history does arb-dex actually have, and for which chains?"

`get_history_summary` sizes the archive *before* you query it:

{
  "rows": 133,
  "rowsWithPairDetail": 103,
  "rowsByEra": { "digest-totals-only": 30, "top-list-pairs": 5, "full-sweep": 98 },
  "pairsTracked": 107,
  "firstAt": "2026-08-10T17:35:09.355Z",
  "lastAt":  "2026-08-15T20:54:34.417Z",
  "spanHours": 123.32,
  "chainsSeen": ["arbitrum", "avalanche", "base", "bsc", "optimism", "polygon"],
  "pairs": [
    { "chain": "polygon",  "pair": "WBTC/USDC", "observations": 103, "qualifiedObservations": 39 },
    { "chain": "bsc",      "pair": "BTCB/USDT", "observations": 99,  "qualifiedObservations": 15 },
    { "chain": "arbitrum", "pair": "ARB/USDC",  "observations": 98,  "qualifiedObservations": 0 }
  ]
}

Coverage is only what was measured. A gap stays a gap — `rowsByEra` says how much detail each

era of rows carries, and `ARB/USDC` having 98 observations but 0 qualified is the archive

telling you that pair has never once cleared the spread floor.

Tools

The two history tools read the service's own measurement archive, so they answer the question

the live tools cannot: whether a dislocation *persisted* or was a single sample. Sampling is

roughly hourly, and gaps are never interpolated or backfilled. Call `get_history_summary` first

to see what span exists before asking for a window.

What it will not do

• Spreads are gross — before gas, MEV and slippage beyond the optimal size. Not a profit

estimate and not trade advice.

• It never fabricates a row. Without a key, `get_chains`, `get_pairs` and `get_spreads`

answer from the free public surface and each carries a `limitation` field naming exactly what

a key would add. `get_prices`, `get_history_summary` and `get_history` return an explicit

key-required error with the signup link rather than a thinner answer dressed up as a full one.

• It does not execute trades, hold funds, or touch a wallet. It is read-only market data.
• This package ships no credentials of any kind. The key is yours and stays in your config.

Get a key

The paid tools call the API through RapidAPI using your own key.

1. Subscribe — there is a free tier:

2. Copy your `X-RapidAPI-Key` from the RapidAPI dashboard.

3. Put it in `RAPIDAPI_KEY` in the config above — never in code, and never in a commit.

Plans

Five of the six tools work on the free tier. Only the per-pair history series is gated:

Calling `get_history` below PRO returns an explicit `tier_required` error naming the plan and

the upgrade URL — it does not fail silently or return an empty series.

Configuration

Set these in your MCP client's `env` block (see the configs above). `.env.example` ships in the

package and documents the same variables for local runs from a clone.

Test

The test suite is not in the npm tarball — run it from a clone:

git clone https://github.com/donnywin85/arb-dex-mcp.git
cd arb-dex-mcp && npm install

npm run selftest                    # keyless: exercises the free fallbacks
RAPIDAPI_KEY=... npm run selftest   # keyed: exercises the paid routes

The test spawns the server over stdio and calls every tool against the real production API —

nothing is mocked. It asserts on live values (block number, per-venue prices, scanned-pair

counts), so a run that passes is evidence the data path works end to end.

Links

• npm:
• Docs:
• Official MCP Registry: `io.github.donnywin85/arb-dex-mcp`
• TensorBlock MCP Index:
• Source:
• The API behind it:

License

MIT — see LICENSE.