Skip to content
DeFiPy

Overview

Large language models are good at reasoning about DeFi positions and poor at computing AMM math. DeFiPy v2 sits in that gap. The 21 analytics primitives shipped in v1 do the math; v2 exposes 10 of them as Model Context Protocol tools, so an LLM can call them the same way a quant calls them from a notebook.

DeFiPy v2 is not an agent framework. It is substrate. There is no orchestration layer, no memory, no planner, no price-oracle service. The library ships three new modules on top of the v1.2.0 primitives:

  • defipy.tools — emits MCP tool schemas for the 10 curated primitives
  • defipy.twin — a state-source abstraction (MockProvider ships; LiveProvider stub is v2.1)
  • a reference MCP server at python/mcp/defipy_mcp_server.py that closes the end-to-end loop

The loop was verified live before release: a Claude Desktop session asked “check the health of the eth_dai_v2 pool,” the MCP server dispatched CheckPoolHealth against a MockProvider-built twin, the primitive ran, the result came back, and Claude correctly inferred the pool’s state and proactively suggested DetectRugSignals as a follow-up. No composition primitive was needed — the curation principle (leaf primitives only, let composition happen LLM-side) held in production.

What this section covers

  • Tool Schemas — how defipy.tools turns a primitive into an MCP tool definition
  • Binding to Claude — wiring the shipped MCP server into Claude Desktop and Claude Code
  • Binding to Other LLMs — the same primitives, other tool-calling interfaces (v2.1)
  • MCP Demo — a walkthrough of the reference MCP server, including the full request/response trace from the verified session

The substrate separation is deliberate. A DeFi analytics primitive that happens to be callable by an LLM is stable; an agent framework that happens to call analytics primitives is not. v2 commits to the first posture.