I’m working on a proxy MCP server that lazy loads tools to save tokens https://github.com/mquan/nimble. It includes a dashboard for connecting and configuring MCP servers.
During the holiday season in December, I got my 9yr old son into programming by making games with LLM. We made a few browser games, and by the end of it he could write the prompt himself, copy the code to the right places, understand what a function does and add objects to the game by modifying an array.
What I learned is llm gives you the activation energy, you just type a few sentences to get the momentum going. To get the reward feedback cycle going you’ll want to add graphics as early as possible.
The most difficult part is to be by his side and ready to jump in whenever a missing coma breaks the entire game. You’ll also need to keep scope under control, I’d alway steer him away from doing any complicated animation. There’re plenty of opportunity to teach but be realistic that it’s not real programming
afaik, the langchain solution loads entire openAPI spec which consumes a lot of token and won't work for many large API. For efficient token usage, api2ai divides the task into two steps: api planning and params parsing. First step takes a summarization of all the endpoints. Once the endpoint is known, we parse params using the schema of the selected endpoint.
I created api2ai (https://github.com/mquan/api2ai) to solve this. You provide an OpenAPI spec and auth data to spin up an agent for your API. It’s still low level and requires dev work. But I’m working on building tools to bring it to end users
I created api2ai, I agree that we want to get consistent result, but that can be solved by looking up cached AI results. If the input is user driven then you’ll need AI to decipher the natural language.
It’s done in two steps:
1. Create an array of every endpoint’s summary text and ask AI to select one based on the user prompt
2. Use function calling with parameters of the selected operation in step 1