langgraph_openai_serve 0.3.3

Openai Compatible Langgraph Server

LangGraph OpenAI Serve

A package that provides an OpenAI-compatible API for LangGraph instances.

Features

• Expose your LangGraph instances through an OpenAI-compatible API
• Register multiple graphs and map them to different model names
• Use with any FastAPI application
• Support for both streaming and non-streaming completions

Installation

# Using uv
uv add langgraph-openai-serve

# Using pip
pip install langgraph-openai-serve

Quick Start

Here's a simple example of how to use LangGraph OpenAI Serve:

from langgraph_openai_serve import LangchainOpenaiApiServe, GraphRegistry, GraphConfig

# Import your LangGraph instances
from your_graphs import simple_graph

async def advanced_graph():
    from langchain_mcp_adapters.client import MultiServerMCPClient
    from langgraph.prebuilt import create_react_agent

    tools = await MultiServerMCPClient().get_tools()
    graph = create_react_agent(model="openai:gpt-4.1", tools=tools)
    return graph

# You can configure your graphs with your desired configurations.
simple_graph_with_history = simple_graph.with_config(
    configurable={"use_history": True},
)
simple_graph_no_history = simple_graph.with_config(
    configurable={"use_history": False},
)

# Create a GraphRegistry
graph_registry = GraphRegistry(
    registry={
        "simple-graph-with-history": GraphConfig(
            graph=simple_graph_with_history, streamable_node_names=["generate"]
        ),
        "simple-graph-no-history": GraphConfig(
            graph=simple_graph_no_history, streamable_node_names=["generate"]
            ),
        "advanced_graph": GraphConfig(graph=advanced_graph, streamable_node_names=["generate"])
    }
)

graph_serve = LangchainOpenaiApiServe(
    graphs=graph_registry,
)

# Bind the OpenAI-compatible endpoints
graph_serve.bind_openai_chat_completion(prefix="/v1")

# Run the app with uvicorn
if __name__ == "__main__":
    import uvicorn
    uvicorn.run(graph_serve.app, host="0.0.0.0", port=8000)

Usage with your own FastAPI app is also supported:

from fastapi import FastAPI
from langgraph_openai_serve import LangchainOpenaiApiServe, GraphRegistry, GraphConfig

# Import your LangGraph instances
from your_graphs import simple_graph, advanced_graph

# Create a FastAPI app
app = FastAPI(
    title="LangGraph OpenAI API",
    version="1.0",
    description="OpenAI API exposing LangGraph agents",
)

# Create a GraphRegistry
graph_registry = GraphRegistry(
    registry={
        "simple_graph": GraphConfig(graph=simple_graph, streamable_node_names=["generate"]),
        "advanced_graph": GraphConfig(graph=advanced_graph, streamable_node_names=["generate"])
    }
)

graph_serve = LangchainOpenaiApiServe(
    app=app,
    graphs=graph_registry,
)

# Bind the OpenAI-compatible endpoints
graph_serve.bind_openai_chat_completion(prefix="/v1")

# Run the app with uvicorn
if __name__ == "__main__":
    import uvicorn
    uvicorn.run(graph_serve.app, host="0.0.0.0", port=8000)

Custom Input, Output, and Context

GraphConfig adapters let a compiled graph use its native LangGraph 1.x input, output, and context schemas. Each adapter can be synchronous or asynchronous and receives the complete chat completion request where applicable.

from dataclasses import dataclass
from typing import TypedDict

from langgraph.graph import StateGraph
from langgraph.runtime import Runtime

from langgraph_openai_serve import GraphConfig


@dataclass
class AppContext:
    user_id: str


class Input(TypedDict):
    question: str


class Output(TypedDict):
    answer: str


class State(TypedDict, total=False):
    question: str
    answer: str


async def generate(state: State, runtime: Runtime[AppContext]) -> Output:
    return {
        "answer": f"Answer for {runtime.context.user_id}: {state['question']}",
    }


graph = (
    StateGraph(
        State,
        input_schema=Input,
        output_schema=Output,
        context_schema=AppContext,
    )
    .add_node("generate", generate)
    .set_entry_point("generate")
    .set_finish_point("generate")
    .compile()
)

graph_config = GraphConfig(
    graph=graph,
    request_to_input=lambda request, messages: {
        "question": messages[-1].content,
    },
    context_factory=lambda request: AppContext(
        user_id=request.user or "anonymous",
    ),
    output_to_text=lambda output: output["answer"],
    streamable_node_names=["generate"],
)

Without adapters, behavior is unchanged: the graph receives {"messages": messages}, no runtime context is supplied, and the response text is read from result["messages"][-1].content. Input and output validation and filtering are performed by LangGraph using the graph's declared schemas.

Using with the OpenAI Client

Once your API is running, you can use any OpenAI-compatible client to interact with it:

from openai import OpenAI

# Create a client pointing to your API
client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="any-value"  # API key is not verified
)

# Use a specific graph by specifying its name as the model
response = client.chat.completions.create(
    model="simple_graph_1",  # This maps to the graph name in your registry
    messages=[
        {"role": "user", "content": "Hello, how can you help me today?"}
    ]
)

print(response.choices[0].message.content)

# You can also use streaming
stream = client.chat.completions.create(
    model="advanced_graph",
    messages=[
        {"role": "user", "content": "Write a short poem about AI."}
    ],
    stream=True
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="")