llm-catalog-pydantic-ai

Pydantic AI adapter for llm-catalog. It turns a role in your catalog config (ai-sdk-catalog.json, shared verbatim with ai-sdk-catalog) into a native Pydantic AI Model — a gateway model is routed through your gateway via the core GatewayTransport, a direct model calls the vendor's own endpoint (anthropic / openai / openai-compatible / google).

import json
from pathlib import Path

from pydantic_ai import Agent
from llm_catalog.pydantic_ai import PydanticAICatalog

config = json.loads(Path("ai-sdk-catalog.json").read_text(encoding="utf-8"))
cat = PydanticAICatalog(config)  # validates the config itself
agent = Agent(cat.model_for_role("fast"))

Each model is built once and reused, with its own HTTP client, so repeated model_for_role() calls share a connection pool. Close the clients with await cat.aclose() when you are done, or use the catalog as async with PydanticAICatalog(config) as cat:.

Installing this package does not pull in litellm. The distributions are deliberately split so a Pydantic AI user's lockfile never references it.

Requires Python 3.12+ and pydantic-ai>=2.51. The HTTP client is httpx2: Pydantic AI and the SDKs it wraps (openai>=3, anthropic>=1) are built on it, and the Anthropic SDK rejects an httpx client outright. The header_rewrite / body_rewrite hooks therefore receive httpx2.Headers / httpx2.Request. Pass transport_factory to supply the underlying httpx2 transport yourself (a proxy, mTLS, connection limits, or httpx2.MockTransport in tests).

See the repository README for the full picture and the verification notes (§9), including the google-genai / custom-HTTP-client caveat.

import namespace: llm_catalog.pydantic_ai