How many times have you copy/pasted from a website into ChatGPT just to ask a quick question about the page? I used to do this constantly with API docs, until this weekend. I realized what I really needed was an AI powered Chrome extension so I can chat with the page directly. So I built one! And it’s powered by a free Pinecone Assistant.

Most AI coding tools tend to fall into a single category, like AI IDEs, AI CLIs, or AI web chat interfaces. The web chats are great for non-developers and make it easy to get started, but lack robust deployment and scaling options. And the CLIs and IDEs are easy to integrate with CI/CD, but harder to set up and require more programming knowledge.

Factory.AI takes a different approach. It’s a web based chat, with a ‘factory bridge’ app running locally to manage files and run terminal commands. The interface is all web based, but the code lives on your machine, making it easy to sync with Git (although you can build in the web if you want).

One feature that really stood out to me was this risk level that shows on every action, along with the running time, and a stop button for that one action.

Claude Code and other AI tools with local file access can be a huge productivity boost, but they can also burn through a lot of tokens! Something I’ve started doing recently is asking Claude to write a function to do X instead of asking Claude to do it directly, anytime I might need to run the same operation again.

For instance, if you’re building an app and need mock data while developing, but a clean install for the repo, you might ask it to clear/delete the mock data, and then generate it again later when testing. Instead, ask Claude to generate the mock data as a CSV and write an import and delete script. Then you’ll only burn tokens once, and you can run the scripts as much as needed.

A few weeks ago at DevRelCon, I sat down with Jim Bennett from Galileo to get his hot takes on AI, and learn how Galileo is tackling the AI hallucination problem. Jim beaks down some of the challenges they have faced in creating evals and observability for AI in customer-facing production applications.

Claude Code is my go-to AI coding tool for most projects. I like to test other AI CLIs, IDEs, and app builders, but I always go back to Claude for anything serious. But occasionally, it will get stuck on a problem and I have to ask another model for advice.

At first, I was copy/pasting from CC to ChatGPT or Gemini, and back to the terminal. But then I set up a Claude Code Hook to call Gemini’s CLI from within Claude Code. I was pretty happy with this solution, but then I saw this even easier approach from Nataly Merezhuk (Nataly Vibe Codes on YouTube).

Claude Desktop runs as an installed app, outside the browser, allowing it to integrate with services running on your local machine, like a database running in Docker. Add an MCP server to that database, and you just gave Claude super powers!

This guide breaks down how to run Neo4j locally with an MCP server connected to Claude Desktop, enabling you to spin up new knowledge graphs from an AI chat with file upload for RAG.

If you enjoy this newsletter, please share it with others in your network. Got a question or comment? Login at news.greenflux.us to share your thoughts.