I Benchmarked CodeAct vs Tool Calling. 91% Fewer Tokens.
The CodeAct paper says LLMs should write code instead of calling tools. I tested it with 120 mock API tools on Mistral Small. The token savings are massive.
Blog
Thoughts on technology, development, engineering, and practical AI workflows
Running Claude Code for Free Locally on Your Laptop
How to run Claude Code locally with Ollama, avoid API costs, install the CLI, choose the right model, and decide when local vs cloud setups make sense.