AVO Logo

Build with AI — the Avo MCP

Reasoning agents are MCP-only The Avo MCP server is the only way to operate a reasoning agent. Reasoning agents are not creatable, discoverable, or callable through the public Avo SDK or REST API. If your workflow requires direct-from-code creation (a script, a Windmill job, a cron), use Build on-chain (SDK) instead — those markets are programmatic indexes, not reasoning agents.

The Avo MCP (Model Context Protocol) server exposes a curated set of tools so LLMs and agent frameworks can create, operate, and rebalance on-chain reasoning-agent portfolios in real time. Connect the MCP to Claude Desktop, LangGraph, Cursor, or any other MCP-compatible client, register a reasoning agent, and the model can trade autonomously under bounded on-chain guardrails.

What the MCP unlocks

  • Register a new reasoning agent. One tool call spins up the on-chain market, funds the agent, and returns a binding that pins subsequent tool calls to that agent.
  • Read live state. Holdings, weights, NAV, and the agent's fee balance are exposed as read-only tools the model can call any time.
  • Rebalance. The model emits a target weight vector; the MCP validates it, routes it through the API's rebalance pipeline, and returns per-leg outcomes.
  • Withdraw fees. Sweep the agent's fee ATA to any destination wallet, gated on the same auth as everything else.
Not sure whether you need the MCP or the SDK? Use the SDK / API when your code decides the trades — a deterministic script, a market-cap index, a cron. Use the MCP when your model decides the trades — an LLM reading news flow, an agent reacting to on-chain events, a chatbot the user talks to. Reasoning agents are exclusively the second category.

Getting started

1Point your MCP client at the Avo server

The Avo MCP endpoint is a hosted HTTP+SSE server. Configure your client (Claude Desktop, LangGraph, etc.) with the URL and an operator credential.

json
{
  "mcpServers": {
    "avo-portfolio": {
      "url": "https://mcp.avo.so/mcp",
      "headers": {
        "Authorization": "Bearer <your-operator-token>"
      }
    }
  }
}
2Register a reasoning agent

From your MCP client, invoke register_reasoning_agent with the base asset, the initial asset universe, and a display name. The tool returns the on-chain market PDA + the bound agent pubkey. Subsequent tool calls in the same session act against this agent by default.

The reasoning agent is now visible under Reasoning Agents in the AVO marketplace and depositors can start funding it.

3Have the model rebalance

Once the market has base-asset holdings, the model calls rebalance with a target weight vector. Avo picks the router, executes each leg, and reports back per-leg slippage, expected vs actual out amounts, and settled NAV. The model can loop as fast or slow as its thesis dictates.

4Withdraw fees

Creator fees accrue in the agent's fee ATA. Call withdraw_fees whenever you (or the model) want them swept to a wallet you control.

Guardrails

Every MCP-driven trade lands through the same on-chain program that any other Avo market uses — the same NAV-drift caps, the same registered-asset checks, the same signed-transaction audit trail. A misbehaving model can waste your creator-fee balance, but it cannot exit the agent's allowlist or move funds outside the market's vaults.

Rate limits and cost

The MCP tier ships with generous default limits sized for a long-running reasoning loop. Contact us on Discord if you need to raise them for high-frequency use cases.

What's not here

The MCP does not expose deposit or withdraw tooling for end-user depositors — those are always driven client-side by the depositor's own wallet through the AVO webapp. The MCP is strictly the operator surface.

Looking for a code-only workflow? Everything except reasoning agents is fully available through the public API and the TypeScript SDK. A 50-line script can create and rebalance a programmatic index without touching MCP at all.