LangGraph API

This package implements the LangGraph API for rapid development and testing. Build and iterate on LangGraph agents with a tight feedback loop. The server is backed by a predominently in-memory data store that is persisted to local disk when the server is restarted.

For production use, see the various deployment options for the LangGraph API, which are backed by a production-grade database.

Installation

Install the langgraph-cli package with the inmem extra. Your CLI version must be no lower than 0.1.55.

pip install -U "langgraph-cli[inmem]"

Quickstart

1. (Optional) Clone a starter template:

   langgraph new --template new-langgraph-project-python ./my-project
   cd my-project
   

   (Recommended) Use a virtual environment and install dependencies:

   python -m venv .venv
   source .venv/bin/activate
   python -m pip install .
   

2. Start the development server:

   langgraph dev --config ./langgraph.json
   

3. The server will launch, opening a browser window with the graph UI. Interact with your graph or make code edits; the server automatically reloads on changes.

Usage

Start the development server:

langgraph dev

Your agent's state (threads, runs, assistants) persists in memory while the server is running - perfect for development and testing. Each run's state is tracked and can be inspected, making it easy to debug and improve your agent's behavior.

A2A wire versions

Send A2A-Version: 1.0 on JSON-RPC requests to opt into v1 response shapes, including streaming responses. V1 responses omit legacy kind discriminators, use statusUpdate and artifactUpdate stream wrappers, emit UTC timestamps ending in Z, and encode file parts with url or raw, mediaType, and filename. Server-generated tool result data uses toolResults in this mode; arbitrary application data and extension payloads are not renamed.

Requests without the header, with an empty header, or with A2A-Version: 0.3, retain the existing response behavior, including tool_results. Surrounding whitespace is stripped, and numeric patch versions are ignored: 1.0.1 selects v1 output and 0.3.1 selects legacy output. Method names alone do not opt into v1 output. Unsupported major/minor versions and malformed header values return JSON-RPC error -32009. All requests accept url and raw file parts in addition to the existing inputs, independent of the response version header. This opt-in covers wire serialization, not full protocol conformance.

A2A task history scope

An A2A context can contain multiple tasks. By default, SendMessage, GetTask, and ListTasks return the full context history. Set historyScope to task to return only messages that belong to the requested task.

For SendMessage, set it in configuration:

{
  "jsonrpc": "2.0",
  "id": "1",
  "method": "SendMessage",
  "params": {
    "message": {
      "role": "ROLE_USER",
      "parts": [{"text": "Second request"}],
      "messageId": "message-2",
      "contextId": "8b1f0e5c-9a3d-4f27-b0c8-2e6a5d4c1b7a"
    },
    "configuration": {"historyScope": "task"}
  }
}

For GetTask and ListTasks, set it directly in params. If changing the request body is not possible, send LangGraph-A2A-History-Scope: task instead. An explicit request value takes precedence over the header.

{
  "jsonrpc": "2.0",
  "id": "2",
  "method": "GetTask",
  "params": {"id": "<taskId>", "historyScope": "task"}
}

{
  "jsonrpc": "2.0",
  "id": "3",
  "method": "ListTasks",
  "params": {"contextId": "<contextId>", "historyScope": "task"}
}

The agent card advertises support through the https://langchain.com/a2a/extensions/history-scope/v1 extension.

How-To

Attaching a debugger

Debug mode lets you attach your IDE's debugger to the LangGraph API server to set breakpoints and step through your code line-by-line.

1. Install debugpy:

   pip install debugpy
   

2. Start the server in debug mode:

   langgraph dev --debug-port 5678
   

3. Configure your IDE:

   • VS Code: Add this launch configuration:

     {
       "name": "Attach to LangGraph",
       "type": "debugpy",
       "request": "attach",
       "connect": {
                "host": "0.0.0.0",
                "port": 5678
            },
     }
     

   • PyCharm: Use "Attach to Process" and select the langgraph process

4. Set breakpoints in your graph code and start debugging.

CLI options

langgraph dev [OPTIONS]
Options:
  --debug-port INTEGER         Enable remote debugging on specified port
  --no-browser                 Skip opening browser on startup
  --n-jobs-per-worker INTEGER  Maximum concurrent jobs per worker process
  --config PATH               Custom configuration file path
  --no-reload                 Disable code hot reloading
  --port INTEGER              HTTP server port (default: 8000)
  --host TEXT                 HTTP server host (default: localhost)

License

This project is licensed under the Elastic License 2.0 - see the LICENSE file for details.