Your expert
coding companion

One of the most exciting capabilities of LLMs like ChatGPT and GPT-4 is their exceptional ability to code. LLM-powered AIs like ChatGPT also make great thought partners in what we should build, and how we should build it.

It’s an exciting time to be making software. With these emerging AI capabilities, a new set of tools for building software are also needed. This is what we’re setting out to deliver with Magnet.

In the future, we believe that AI coding assistants (agents) will be able to go end-to-end on shipping a wide variety of features and fixing most errors that happen in your software. From prioritizing features, to designing them, to implementing them in code. These AI coding assistants will substantially speed up existing software developers and also make coding accessible to a much wider audience. The time it takes to go from having an idea, to fleshing it out and getting to a working piece of software will become shorter and shorter in the coming years.

The initial focus we’ve had when building Magnet has been exploring interfaces that give humans the best controls for leveraging AI coding superpowers - emphasizing the unique collaboration between humans and AI to build software.

The vision for Magnet in the coming months is to build out:

• A truly integrated coding assistant, that is aware of all the aspects of the software development process, and all the tools you use to do your work - the thinking here is that with more and more context, AI will be able to do a better and better job at helping you build.
• We’re not looking to build another IDE, but instead want to reimagine interfaces leveraging the latest capabilities of AI - we believe things can look a lot different.
• We want this AI companion to act as a thought-partner, and a super-charged, super-knowledgable coding assistant that implement features end-to-end.

The alpha version we’re shipping today, is a preview of that longer-term vision - but we already find it useful and think you might as well. We’ve been using Magnet to build Magnet, and this bootstrapping loop has been fun.

If you find yourself copying and pasting snippets of code from your codebase into ChatGPT, then this first version of Magnet is for you. Often, when turning to ChatGPT for help, you’ve already got a codebase you’re working on, but ChatGPT doesn’t have access to that code. We found ourselves often copying and pasting the content of files into the ChatGPT UI in order to help give ChatGPT sufficient context on the code we’re working with. This copy and paste process to give ChatGPT context on our codebase felt tedious, and so we wanted to make this a much smoother process.

The first version of Magnet solves this UX challenge. With Magnet, you can open any directory you’re working on locally and:

• Provide multi-file coding context effortlessly. When working on a new feature, you can select multiple files that you know are relevant to that feature, and prompt Magnet with the context of all of those files. Magnet will make intelligent, contextual recommendations for how to modify your code to deliver what you’re working on.
• Powerful GPT-4 code suggestions. By default, Magnet is free if you use your own OpenAI API key, and we support both GPT-3.5 and GPT-4 API keys. If you don’t yet have access to a GPT-4 API key, or you don’t want to worry about API keys in general, sign up for the Pro plan to get access to a GPT-4 powered coding assistant powered by our hosted service.

We’d love to get your feedback on this early version! Join us on Discord.

A few things on our roadmap include:

• Issues integrations. Integrations with Linear and Github Issues so you can start your Magnet work from an existing ticket.
• Relevant context suggestions. Magnet suggestions on which parts of the code might be relevant to the task at hand, to make selecting the right parts of the codebase even easier.
• Change suggestions and pull requests. Magnet will create an end-to-end set of changes that it suggests you implement, and can also open these changes as a separate pull request.
• Tool-based agentic coding. Assistants that can ship end-to-end features starting from a simple description.