Claude Coder Agent

Inspiration

Learning new library is always hard, reading blog post can be outdated. Why not learn by running?

What it does

Our agent can help you learn by reading, memorizing and executing the code from any GitHub repo 🧙

How we built it

Our agent uses the powerful Claude 2 LLM, and ChromaDB as its memory, to embed any github repo that you are interested in. It can then run the code it generates in an isolated container environment, it uses the output or error to iterate until it has a perfect answer for you!

Challenges we ran into

Langchain agent does not work with Claude 2 out-of-box, we used prompt engineering that we learned during the hackathon to steer the agent to know how to use the tools and give us the output we need.

Accomplishments that we're proud of

We are able to let the agent combine multiple tools: code_search and code_interpreter.

It can search across the vector database to retrieve the most up to date information
It can then synthesize the code from the information it gathered to execute the code in a safe, isolated environment.
It uses the code output or error to iterate its thought process, until it have a final answer for the user.

What we learned

Building a safe, useful AI application that integrates with real-world knowledge base, and being able to execute arbitrary code safely.

What's next for Claude Coder Agent

Use it to learn more library/framework! And productionize claude-coder so others can benefit. We believe the next generation of learning is explorative and interactive.