tool-server-openapi-spec

OpenAPI 3.0.0 spec for an Aptible tool server. A tool server exposes custom tools to Aptible AI.

The specs in this repository can generate conformant scaffolding for a tool server.

Getting Started

Install openapi-generator and choose a language/framework from the list of supported OpenAPI server stubs.

For the rest of this example, we'll assume Python with the Flask framework for concreteness and walk through the steps to implement a "Hello, World!" tool server. Server generation in other frameworks will involve similar steps.

First, in a directory where you want to implement the server, generate the scaffolding:

openapi-generator generate \
-i https://raw.githubusercontent.com/aptible/tool-server-openapi-spec/main/1.0.0/tool-server.yaml \
-g python-flask

You now need to fill in the implementation of the controllers. First, edit openapi_server/controllers/security_controller.py to make your server validate the bearer authentication token that Aptible will pass for authentication. A basic implementation that assumes you're passing the token into the server as an environment variable requires filling in a few lines to complete the existing info_from_bearerAuth stub:

import hmac
import os


def info_from_bearerAuth(token):
    if hmac.compare_digest(token, os.environ['TOOL_SERVER_TOKEN']):
        return {}  # Success
    return None  # Failure

Next, edit openapi_server/controllers/security_controller.py to add an implementation of the get_tools function:

from openapi_server.models.tool import Tool
from openapi_server.models.tools_list import ToolsList

def get_tools():
    return ToolsList(
        [
            Tool(
                tool_id="hello_world",
                description="Returns the string \"Hello, World!\" when you pass JSON input { \"name\": \"World\" }."
                            "Pass a different name for a different greeting.",
                human_short_title="Hello World Generator",
                human_description="Generate a personalized greeting by passing a JSON object with a \"name\" key."
            )
        ]
    )

and the call_tool function:

from openapi_server.models.tool_call_result import ToolCallResult

def call_tool(tool_id, body):
    if tool_id == "hello_world":
        name = body['name']
        return ToolCallResult({"summary": f"Hello, {name}!"})
    else:
        raise NotFound(f"No tool registered with ID {tool_id}")

Follow instructions in the generated README.md to install dependencies, set a TOOL_SERVER_TOKEN for testing in your shell environment, and start your server. If you're testing locally, you should now be able to hit the tool server endpoints and see everything working:

$ curl http://localhost:8080/tools -H "Authorization: Bearer ${TOOL_SERVER_TOKEN}"
{
  "tools": [
    {
      "description": "Returns the string \"Hello, World!\" when you pass JSON input { \"name\": \"World\" }. Pass a different name for a different greeting.",
      "human_description": "Generate a personalized greeting by passing a JSON object with a \"name\" key.",
      "human_short_title": "Hello World Generator",
      "tool_id": "hello_world"
    }
  ]
}

$ curl -X POST http://localhost:8080/tool/hello_world -H "Authorization: Bearer ${TOOL_SERVER_TOKEN}" -H "Content-Type: application/json" -d '{"name": "World"}'
{
  "result": {
    "summary": "Hello, World!"
  }
}

Now add some real tools and run the server behind a HTTPS endpoint to serve your tools in production.