Binding to Other LLMs

v2.0 ships MCP schemas only. Anthropic tool-use and OpenAI function-calling adapters are deferred to v2.1 and will be derivable from the shipped MCP schemas with small wrappers.

defipy.tools.get_schemas(format="anthropic") and get_schemas(format="openai") will raise NotImplementedError in v2.0. The MCP format is strategically first for v2.0 because MCP is the direction the agent ecosystem is heading — Claude Desktop, Claude Code, and third-party MCP clients all consume it natively. Shipping DeFiPy as MCP-native is stronger positioning than emitting vendor-specific tool-use JSON.

A developer who needs an Anthropic or OpenAI schema today can derive one from the shipped MCP schema directly. The input-schema surface is nearly identical across formats — it’s the envelope (description placement, required fields, tool-routing semantics) that differs. For users with an urgent need, a ~30-line adapter script is straightforward.

What v2.1 will ship

• get_schemas(format="anthropic") — returns the list of schemas in Anthropic tool-use JSON shape
• get_schemas(format="openai") — returns the list in OpenAI function-calling shape
• Reference examples for both in python/examples/

Workarounds for today

1. Roll a derivation script. Load the MCP schemas via get_schemas("mcp"), transform the envelope. The input_schema field moves to input_schema (Anthropic) or parameters (OpenAI); the name and description fields transfer unchanged.
2. Wait for v2.1. Timeline tracked in the Roadmap.

Note

The primitives themselves are LLM-agnostic. Whatever format wraps them, the .apply() call is the same code path. That’s the architectural commitment — the primitive is the tool, regardless of which vendor’s tool-use format is in play.