langchain-pi 0.4.0

LangChain BaseChatModel adapter for Pi (@earendil-works/pi-ai) via a Node sidecar — use Pi with any provider, model and auth from LangChain/LangGraph in Python. Includes a native opencode/Zen chat model (no pi).

langchain-pi

A LangChain BaseChatModel adapter for Pi (@earendil-works/pi-ai), the Python twin of langchain-pi-ts. Use Pi — and any provider, model and credential it resolves — from LangChain and LangGraph in Python, with native tool calling and streaming.

Pi is a TypeScript/Node stack with no Python SDK, so langchain-pi drives a tiny Node sidecar that wraps pi-ai's streamSimple + ModelRegistry/AuthStorage. Node keeps owning provider dispatch, OAuth refresh and the provider stealth headers; Python only frames the request and reconstructs the streamed events. Python never touches a token.

Requirements

• Python ≥ 3.9
• Node ≥ 22.19.0 on PATH, with @earendil-works/pi-ai and @earendil-works/pi-coding-agent resolvable (install them where the sidecar can reach them, or point node_modules_dir at a node_modules directory that has them). Note: NODE_PATH does not work for ESM, so node_modules_dir resolves each package via its package.json entry.

Install

pip install langchain-pi          # or: uv add langchain-pi
npm install @earendil-works/pi-ai @earendil-works/pi-coding-agent

Usage

create_chat is the simplest entry — pass a provider and it routes to the right backend (opencode/opencode-go → native Zen, everything else → Pi):

from langchain_pi import create_chat

free = create_chat("opencode", "deepseek-v4-flash-free")          # free, no key
go = create_chat("opencode-go", "glm-5", api_key="...")           # subscription
codex = create_chat("openai-codex", "gpt-5.3-codex-spark")        # via Pi

Or construct a backend directly:

from langchain_pi import ChatPi

model = ChatPi(
    provider="openai-codex",
    model="gpt-5.3-codex-spark",
    reasoning="minimal",
    system="You are a helpful assistant.",
)

print(model.invoke("Hello!").content)

Tool calling

Accepts any LangChain tool; schemas are converted to the JSON Schema pi-ai expects.

from langchain_core.tools import tool

@tool
def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    return f"It is sunny in {city}, 24C."

agent_model = model.bind_tools([get_weather])

Streaming

for chunk in model.stream("Write a haiku."):
    print(chunk.content, end="")

LangGraph

from langgraph.prebuilt import create_react_agent

agent = create_react_agent(model, [get_weather])
agent.invoke({"messages": [("user", "What's the weather in Paris?")]})

Pointing at your Node install

If pi-ai isn't resolvable from the sidecar's location, pass the node_modules directory that contains it:

ChatPi(provider="openai-codex", model="gpt-5.3-codex-spark",
       node_modules_dir="/path/to/your/project/node_modules")

Native opencode/Zen (no Pi)

OpenCode Zen is an OpenAI-compatible endpoint, so you don't need Pi. langchain_pi.opencode ships ChatOpencode, a thin ChatOpenAI subclass pointed at it (langchain-openai ships as a dependency).

The API key is optional: free models work with no key (anonymous, IP-rate- limited). For paid models pass a key via OPENCODE_API_KEY (env) or api_key.

from langchain_pi.opencode import ChatOpencode

free = ChatOpencode("deepseek-v4-flash-free")  # no key
paid = ChatOpencode("glm-5")  # OPENCODE_API_KEY or api_key
go = ChatOpencode("glm-5", tier="go")

Free models: deepseek-v4-flash-free, big-pickle, mimo-v2.5-free, nemotron-3-super-free.

Notes

• Node is a runtime prerequisite — the heavy provider/auth logic lives in pi-ai.
• Tool-call deltas and usage/cost metadata are reconstructed 1:1 with the TS twin.

License

MIT