open-langchain

Native LangChain chat models for OpenAI Codex (ChatGPT Plus/Pro subscription OAuth), Claude (Claude Code subscription OAuth), and OpenCode Zen/Go. No Pi runtime, no Node sidecar.

Requirements

• Python >= 3.9
• For Codex: an openai-codex credential in ~/.pi/agent/auth.json, or sign in with codex-login
• For Claude Code: the Claude Code CLI logged in once (~/.claude/.credentials.json)
• For paid OpenCode models: OPENCODE_API_KEY or an explicit api_key

Install

pip install open-langchain
# or: uv add open-langchain

Usage

create_chat routes the supported native providers:

from open_langchain import create_chat

codex = create_chat("openai-codex", "gpt-5.3-codex-spark")
claude = create_chat("claude-code", "claude-sonnet-4-6")
free = create_chat("opencode", "nemotron-3.5-lightning-free")
go = create_chat("opencode-go", "minimax-m3", api_key="...")

Or construct Codex directly:

from open_langchain import ChatCodex

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

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

Codex models: gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-6-astra, gpt-5.5, gpt-5.3-codex-spark. As of 2026-09 the ChatGPT account only serves these; older GPT-5.x ids return 400 not supported when using Codex with a ChatGPT account. calculate_cost follows models.dev pricing, including the long-context tier that bills the whole request at the higher rate once the prompt exceeds 272k tokens. Ids outside the catalog are still sent to the API, but calculate_cost prices them as unknown and returns 0.

Codex Auth

ChatCodex reads the same ~/.pi/agent/auth.json credential shape under openai-codex, refreshes the OAuth token in place, and talks directly to https://chatgpt.com/backend-api/codex/responses.

If no credential exists:

codex-login
codex-login --device

Tool Calling

from langchain_core.tools import tool
from open_langchain import ChatCodex

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

model = ChatCodex(model="gpt-5.5").bind_tools([get_weather])
msg = model.invoke("What's the weather in Paris?")
print(msg.tool_calls)

tool_choice is passed through to the Codex Responses API:

forced = ChatCodex(model="gpt-5.5").bind_tools(
    [get_weather],
    tool_choice={"type": "function", "name": "get_weather"},
)

For agent loops, keep tool_choice="auto" unless every model turn should call the same tool.

Streaming

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

OpenCode

ChatOpencode uses OpenCode's OpenAI-compatible endpoints through langchain-openai.

from open_langchain import ChatOpencode

free = ChatOpencode("nemotron-3.5-lightning-free")
paid = ChatOpencode("minimax-m3", api_key="...")
go = ChatOpencode("minimax-m3", tier="go", api_key="...")

Free models include nemotron-3.5-lightning-free, big-pickle, and mimo-v2.5-free.

Claude Code (Anthropic subscription)

ChatClaudeCode talks to the Anthropic Messages API authenticated with the Claude Code OAuth session already on the machine (~/.claude/.credentials.json), billing requests against your Claude Code subscription — no API key. It reads and refreshes the token in place (with a claude CLI fallback). On macOS it reads the login keychain item Claude Code-credentials first and falls back to ~/.claude/.credentials.json, writing the refreshed token back to whichever source it read. An explicit creds_path always wins and uses only that file.

from open_langchain import ChatClaudeCode, create_chat

chat = create_chat("claude-code", "claude-sonnet-4-6")
print(chat.invoke("Hello!").content)

# Or construct directly, with options:
opus = ChatClaudeCode(model="claude-opus-5", reasoning="medium")

Models: claude-opus-5, claude-fable-5-1, claude-fable-5, claude-opus-4-8, claude-opus-4-7, claude-sonnet-5, claude-sonnet-4-6, claude-haiku-4-5. Reasoning uses adaptive thinking on the Claude 5 family and Opus 4.8/4.7, and a token budget on Sonnet 4.6 (Haiku has no reasoning). The Claude 5 models ship a 1M context window by default; for the older models the 1M-context beta is opt-in via long_context=True, since the subscription rejects long-context requests without extra credits otherwise. Tool calling and streaming work as with ChatCodex.

tool_choice is forwarded to the Messages API unchanged. claude-fable-5-1 rejects a forced tool choice - "any" or a named tool - with a 400, so use tool_choice="auto" with that model.

Using a subscription OAuth session from a third-party app may violate Anthropic's terms and risk your account. See the pi-claude-code-auth README before relying on this.

License

MIT