Azure Functions LangGraph

Deploy existing LangGraph graphs as Azure Functions Python v2 HTTP endpoints without writing request wrappers. This is a deployment adapter, not an agent builder or a replacement for LangGraph.

한국어 · 日本語 · 简体中文

Installation

Python 3.11–3.14 is supported. Azure Functions hosting support depends on the selected plan/runtime; see deployment.

pip install azure-functions-langgraph

Quick Start

Save as function_app.py in a Functions app with host.json:

from typing import TypedDict
from langgraph.graph import START, END, StateGraph
from azure_functions_langgraph import LangGraphApp
class State(TypedDict):
    message: str
builder = StateGraph(State)
builder.add_node("echo", lambda state: {"message": state["message"]})
builder.add_edge(START, "echo")
builder.add_edge("echo", END)
app = LangGraphApp()
app.register(builder.compile(), "echo")
func_app = app.function_app

Run func start, then POST {"input":{"message":"hello"}} to http://localhost:7071/api/graphs/echo/invoke; the result is {"output":{"message":"hello"}}. Deployed routes require a function key by default. Follow the 5-minute tutorial for complete project files, local setup and deployment.

Deploy to Azure (default path)

One path, start to finish: the 5-minute tutorial: local invoke, minimal Azure resources, publish, authenticated cloud invoke, cleanup. Stateless graphs need only the Storage Account Azure Functions itself requires. Premium, Blob/Table persistence, Platform compatibility, Managed Identity and distributed locks are optional next steps in the advanced deployment guide. Cloud steps are not claimed as certified beyond the stability table below.

Stability — rolling 0.x, per surface

The package-level classifier is Beta, not a promise that every surface is production-ready. Compatibility decisions follow CONTRIBUTING.md; breaking changes remain on the 0.x line under the existing release policy.

Classic stream routes return buffered SSE, not incremental tokens. StreamingLangGraphApp is a separate experimental transport. Function keys are not user/thread ownership; applications must enforce authorization. Default locks and Platform thread records are process-local. Shared writers need explicit distributed storage/locking; see the production guide and security boundaries.

Recommended learning path

Optional, after the default path. These are independent, capability-focused apps, not cumulative production templates or prerequisites:

1. Azure OpenAI agent
2. Conversation memory
3. Tool calling
4. Persistent production wiring

Also optional: an official LangGraph guide graph hosted with minimal changes, an official-example adaptation (not one of the custom examples above).

Examples support deterministic fake models without cloud credentials. For Platform thread storage, the tested Managed Identity storage recipe wires Azure Table explicitly. State separation is not authenticated ownership.

Documentation

• Configuration and usage
• Adapter reference, optional features and migrations
• Architecture and distributed deployment
• Finite SDK support matrix: partial compatibility, not Platform parity
• Examples, troubleshooting, security
• Contributing, product scope, design
• AI tooling: llms.txt, llms-full.txt, AGENTS.md

Contributing

Contributions are welcome. The Contributing Guide covers local setup, the checks to run, and the pull request workflow.

Looking for something to work on?

• Good first issue: small, self-contained starter tasks.
• Help wanted: tasks that may need more familiarity with this codebase or Azure Functions.

Please comment on an issue to request assignment before you open a pull request. Either list can be empty at any time, so bug reports and small documentation suggestions are always welcome too.

Part of the Azure Functions Python DX Toolkit. Integrated application recipes live in the cookbook; this package owns serving compiled graphs, not prompts, models or tool frameworks.

Independent community project; not affiliated with Microsoft or LangChain. Azure, Azure Functions and LangGraph are their owners' trademarks. MIT.