Skip to content

Latest commit

 

History

History
60 lines (35 loc) · 3.24 KB

README.md

File metadata and controls

60 lines (35 loc) · 3.24 KB
Raw

MoonBit AI Agent using Components

Requirements

The following tools are needed:

and of course, moonbit toolchain.

For this example, you will also need to create a GitHub App and obtain the private key, the client ID and the App ID and the secret. You may consider following the step 2 through 5 with the official tutorial. You may configure the callback URL of the GitHub App with the Spin application for production or a proxy (e.g. <smee.io>) for local development.

You also need to have access to an AI provider, such as DeepSeek or Silicon Flow. You need an API Key to access the LLM, the endpoint the the completion functionality, and the name of the LLM that you'd like to use. We are using the DeepSeek Coder provided by Silicon Flow in this demo.

How to use

  • Copy the folder gen to the one you like, and then implement the let server global variable.
  • Add configurations in spin.toml as the existing one.

Build commands:

  • Install dependency with moon update && moon install
  • Local development: spin build && spin up.
  • Publish: ./publish.sh. Don't forget to spin login the service.

Environment variables

Copy .env.template to .env and fill in the environment variables.

They are configured in variables for spin.toml and component.agent.variables will use them so that component can have access. The example application use the environment variables to pass secrets. One may use external vault in production. Check out the official document.

Project Structure

Most files are generated with

wit-bindgen moonbit --derive-show --derive-eq --derive-error --out-dir . wit

Note that the MoonBit support is accepted by the bytecode-alliance/wit-bindgen so there is where you should install the wit-bindgennow.

Apart from the auto-generated code:

  • agent: an example of AI Agent using MoonBit.

  • http: a wrapper around the underlying wasi-http interface, providing proper lifetime handling. One should always call response.body.trailer before it goes out of scope. It is built with the concept of "Your Server as a Function" as http4k as well as the wasi-http itself.

  • crypto: containing a jwt RS256 implementation for signing GitHub access token, and necessary RSA / HMAC algorithms.

  • io: a thin wrapper around the underlying wasi-io streams.

  • utils: some helper functions for encoding/decoding base64 and utf-8.

Future Work

As of the project itself, there are several things we can improve:

  • We should validate incoming deliveries with the HMAC algorithms (which is implemented)
  • Use Vault for keys
  • Use Redis Trigger for long running task (though not yet supported by fermyon cloud)