graphserve 0.1.1

Serve any LangGraph graph over the OpenAI Chat Completions + Responses APIs.

GraphServe

GraphServe is an open-source Python library that serves any LangGraph graph over the OpenAI Chat Completions and Responses APIs, enabling seamless integration of agent workflows into OpenAI-compatible applications.

Installation

pip install graphserve

Quickstart

from fastapi import FastAPI
from langgraph.graph import StateGraph, START, END
from graphserve import GraphRegistry, GraphConfig, create_openai_router

# 1. Build your LangGraph graph
graph = StateGraph(...)
# ... add nodes and edges ...
compiled = graph.compile()

# 2. Register it under a model name
registry = GraphRegistry()
registry.register("my-agent", GraphConfig(graph=compiled))

# 3. Mount the OpenAI-compatible router on your FastAPI app
app = FastAPI()
app.include_router(create_openai_router(registry), prefix="/v1")

Your app now exposes:

Route	Description
GET /v1/models	List registered graphs
POST /v1/responses	Create a response (streaming or non-streaming)
GET /v1/responses/{id}	Retrieve a previous response
DELETE /v1/responses/{id}	Delete a response
POST /v1/chat/completions	Chat Completions API

Public API

Export	Description
GraphRegistry	Registry mapping model names to graph configs
GraphConfig	Holds a single already-compiled graph to serve
create_openai_router	Builds a FastAPI APIRouter with all OpenAI-compatible routes

create_openai_router options

create_openai_router(
    registry,   # GraphRegistry — required
)

GraphServe is a pure OpenAI↔LangGraph converter. Cross-cutting concerns are the consumer's job, applied with standard tools:

• Auth — apply it where you mount the router:

  app.include_router(create_openai_router(registry), prefix="/v1",
                     dependencies=[Depends(verify_api_key)])
  

• Callbacks / tracing — attach to your graph when you build it.

Per-request runtime context is derived generically from the OpenAI request and exposed to the graph as LangGraph runtime context: user → context["user_id"], instructions → context["metadata"]["custom_instructions"], and metadata is passed through as context["metadata"]. Graphs read what they need and ignore the rest.

Stateful responses and previous_response_id

All state is managed through the registered graph's LangGraph checkpointer, keyed by thread_id. If a graph is compiled without a checkpointer, GraphServe automatically injects an InMemorySaver (with a warning). Response IDs encode the model (resp_<model>.<hex>) so GET/DELETE routes can resolve the owning graph without any external metadata store.

To use a persistent checkpointer:

from langgraph.checkpoint.memory import MemorySaver
compiled = graph.compile(checkpointer=MemorySaver())

Streaming

Pass "stream": true in the request body to receive Server-Sent Events following the OpenAI Responses API event schema (response.created, response.output_item.added, response.output_text.delta, response.completed, etc.).

License

MIT