A Crypto Portfolio MCP Server for AI Agents

An AI agent asked to analyse a wallet is only as good as the data it can reach. Give it a token-only API and it will produce a fluent, well-structured, badly wrong answer — because it has no way to know that the lending, staking and perp positions were never in the response.

The Model Context Protocol is how you hand an agent a real data source. Octav ships an MCP server for exactly this.

What the MCP server exposes

Once connected, the agent can call the Octav API directly as tools — portfolio, transactions, NAV, historical snapshots and the rest of the endpoint surface — without you writing any glue code.

Setup

The server is at github.com/Octav-Labs/octav-api-mcp. Add it to your MCP client configuration:

{
  "mcpServers": {
    "octav": {
      "command": "npx",
      "args": ["-y", "@octav/api-mcp"],
      "env": { "OCTAV_API_KEY": "your-key" }
    }
  }
}

Then ask in natural language:

What DeFi positions does 0xd8dA6BF26964aF9D7eEd9e03E53415D37aA96045 hold,
and what is the total exposure by chain?

The agent calls the portfolio tool, receives decoded positions, and reasons over real data instead of guessing.

Why decoding matters more for agents than for humans

A human looking at a dashboard notices when a position is missing — they know they have a Kamino loan and can see it is absent. An agent has no prior. It treats the API response as the complete state of the world.

That makes silent omission uniquely dangerous in agent workflows. An API that returns 200 OK with half a portfolio does not produce a visible error; it produces a confident conclusion built on missing data. The mechanics of that failure are covered in Why Portfolio APIs Disagree About Net Worth.

The rest of the agent toolkit

The MCP server is one of four ways to give an agent portfolio access:

A note on rate limits and cost

Agents are enthusiastic. They will happily call an endpoint in a loop.

The API allows 360 requests per minute per key and bills in credits at roughly $0.020–$0.025 each, with most endpoints costing one credit. Two practical defences: pass multiple addresses in a single request rather than looping (?addresses=0x123,0x456), and rely on the built-in caching — portfolio responses cache for one minute, transactions for ten.

If you want hard spend limits per call rather than a shared key, that is what x402 is for.