AutoPlugin

AutoPlugin is a Python package that makes it easy to convert Python functions into ChatGPT plugins. With just a couple lines of code, you can:

• Automatically create an OpenAPI spec with custom endpoints for your registered Python functions, telling ChatGPT how to use it. Pull endpoint descriptions from the function docstring or generate them automatically with the OpenAI API.
• Generate the ai-plugin.json file to register your plugin with ChatGPT.
• Launch a local server that can be used by ChatGPT for development.

Installation

To install AutoPlugin, simply run the following command:

pip install autoplugin

To install with the ability to generate endpoint descriptions for the OpenAPI specification automatically from source code, install with

pip install 'autoplugin[gen]'

Basic Usage

To get started with AutoPlugin, follow these steps:

1. Import the necessary functions from AutoPlugin:

from autoplugin import register, generate, launch, get_app

2. Create an app instance, backed by FastAPI:

app = get_app()

3. Use the register decorator to register your functions as API endpoints. AutoPlugin will automatically generate descriptions if needed.

@register(app, methods=["GET"])
async def get_order(name: str) -> str:
    order = await get_order_from_db(name)
    return f"Order for {name}: {order}"
# Generated description: "Retrieves an order from the database for a given name."

4. Generate the necessary files (openapi.yaml and ai-plugin.json) for your ChatGPT plugin. Optionally, specify out_dir to change where they're saved to, or set overwrite_openapi_spec=False or overwrite_plugin_spec=False to avoid overwriting the respective files.

# generated files saved to `.well-known/` directory
generate(app, name="Example", description="Plugin to add numbers or greet users")

5. Launch the server. Optionally, specify host and port:

launch(app)  # API hosted at localhost:8000

6. Follow the instructions to run a custom plugin:

• On ChatGPT, make a new chat.
• Under "Models" select "Plugins"
• In the Plugins dropdown, select "Plugin store"
• Click "Develop your own plugin"
• Enter the URL you're running the server at ("localhost:8000" by default) and hit enter.
• Click "Install localhost plugin"

Example

Here's a complete example that demonstrates how to use AutoPlugin to create API endpoints for two functions, hello and add. It also generates the openapi.yaml and ai-plugin.json files, by default in the .well-known directory. :

from autoplugin.autoplugin import register, generate, launch, get_app

app = get_app()

@register(app, methods=["GET", "POST"])
async def hello(name: str, age: int = 5) -> str:
    return f"Hello, {name}! Age {age}."

@register(app, methods=["GET"])
async def add(a: int, b: int) -> int:
    """ Adds two numbers """
    return a + b


# Generate the necessary files
generate(app, name="Example", description="Plugin to add numbers or greet users")

# Launch the server
launch(app)

This example creates a FastAPI server with two endpoints, /hello and /add, that can be accessed using GET or POST requests. AutoPlugin will use the docstring for the OpenAPI description of /add and generate an automatic description for /hello by passing the source code of the function to OpenAI's API.

Docs

The @register Decorator

The @register decorator is used as follows:

@register(app: FastAPI,
            methods: List[str],                     # which HTTP methods to support
            description: Optional[str],             # if provided, used as is
            generate_description: Optional[bool])   # whether to autogenerate a description
def my_func(...):
    ...

AutoPlugin generates function descriptions in the OpenAPI spec so that ChatGPT knows how to use your endpoints. There are a few arguments to customize the behavior of this generation.

• app: Your FastAPI application. AutoPlugin provides a get_app function that includes CORSMiddleware for testing convenience (allows all origins by default).
• methods: A list of HTTP methods to be supported (e.g. ”GET”, POST”)
• description: If provided, overrides everything else and is used directly as the endpoint description for the OpenAPI spec
• generate_description: If set to True, AutoPlugin will generate one automatically from OpenAI's API (requires the LangChain package and setting the OPENAI_API_KEY environment variable).

By default (if neither description nor generate_description are provided), the description is fetched from the docstring. If there's no docstring, AutoPlugin falls back to generating one automatically.

The generate Function

The generate function has the following signature:

def generate(app: FastAPI, version="v1", out_dir=".well-known",
             overwrite_plugin_spec=True, overwrite_openapi_spec=True,
             name="", description="",
             **kwargs)

• app: Your FastAPI application again.
• version="v1": What version number to pass to both the plugin and OpenAPI specs.
• out_dir=".well-known": The directory to save both files to.
• overwrite_plugin_spec=True: If set to False, does not overwrite ai-plugin.json if it already exists.
• overwrite_openapi_spec=True: If set to False, does not overwrite openapi.yaml if it already exists.
• name="": If specified, used for both name_for_human and name_for_model.
• description="": If specified, used for both description_for_human and description_for_model. Keep in mind the best practices for descriptions.
• **kwargs: All other keyword arguments are passed on to ai-plugin.json directly. See the full list of possible options here.

The launch Function

The launch function has the following signature:

def launch(app: FastAPI, host="127.0.0.1", port=8000):

• app: Still your FastAPI application.
• host="127.0.0.1": the host to launch the server on
• port=8000: the port to launch the server on

Testing

AutoPlugin also provides a testing_server utility (courtesy of florimondmanca) for testing your endpoints. Here's an example of how you can use it to test the /hello and /add endpoints from the example above:

from autoplugin.testing import testing_server
from os.path import join
import requests

def test_api():
    host = "127.0.0.1"
    port = 8000
    server, base_url = testing_server(host=host, port=port, app_file="path/to/example.py", app_var="app")

    with server.run_in_thread():
        # Server is started. Do your tests here.
        response = requests.post(join(base_url, "hello"), json={"name": "John Doe", "age": 31})
        assert response.json() == {"result": "Hello, John Doe! Age 31."}

        response = requests.get(join(base_url, "hello"), params={"name": "Jane Smith"})
        assert response.json() == {"result": "Hello, Jane Smith! Age 5."}

        response = requests.get(join(base_url, "add"), params={"a": 6, "b": 8})
        assert response.json() == {"result": 14}
        # Server will be stopped.

test_api()