I’m looking for guidance on how to start using AI for Julia coding and in general.
So far, my AI experience is asking questions in new Google search tabs. I tried searching for articles and videos on how to get started, but they all feel like ads. The projects I’ve seen announced to “make it easy” to use AI with Julia (e.g. SpaceStation, Kaimon, BonitoBook, AIHelpMe) have differing setup instructions that I don’t understand, requiring some combination of servers, secrets, keys, configs, gates, JSON files, prepaid credits …
Can someone recommend a simple and preferably free way to dip my toe in?
The most basic setup I can think of (but that can take you much further that you might think) is an harness like pi (there are many others, but this one is very minimal) and a provider. Then you connect the harness to the provider via an API key or a simple login and you’re good to go.
Regarding the “Free” part it’s harder. If you have the hardware you can run models locally (small ones don’t need a lot of memory), but that is more of a “head dive” rather than “dipping the toes”. Otherwise, you can go either subscription or consume based where you top up an amount.
I personally am liking a lot the opencode-go subscription, which is 10eur with decent limits and decent model selection. Sometimes you can find models that are listed as free, but it goes without saying that the payment in that case is the data you provide for training (also usually heavily throttled).
Overall in terms of providers, you can go directly to the mothership: OpenAI, Anthropic,Z.AI,Deepseek,etc.. or go for “aggregators”, like opencode-go or openrouter.
The tradeoff is, more model selection, but usually not as cheap as if going directly to the tap.
There are many providers that also offer various levels of data retention policies and or training, especially when talking about open weight models.
I have zero experience with notebooks, but I can share how I use VS Code for programming as well as writing boring documents. I do a bit of hobby coding, and a bit for work, but it is not 5 days 8 hours a week.
I have a Google One AI Premium subscription, costing about 20€$. With that, I use Gemini heavily for all kinds of questions, privately and professionally. Up to six family members can be included under one plan, BTW. In VS Code, I installed the Google Antigravity extension. For the price, it provides relatively generous access to Gemini agents, and a bit of Claude, which is definitely more intelligent. Next, I installed Kaimon according to the setup instructions.
Once configured, you can open your workspace and start AI-assisted coding. For configuration questions, either the built-in Gemini agent or the web version can assist. Ask them also to help you with creating the skills.
Drafting routine documents follows the same basic workflow: feed relevant information into the workspace or reference specific URLs, and let the agents generate a Markdown file. From there, you edit it manually depending on the intended use.
Keep in mind that sensitive data does not belong in agent workflows: Weigh the risks.
Bite the bullet and sign up for a $20/month subscription to Claude. It’s the best all-around (And, no, I don’t get a commission.) This link has Claude’s own from scratch setup, including installation of Julia for completeness.
Top gotcha: Keep your sessions short to avoid context rot because quality falls off over time, and you will see it doubling down on dead ends.
Practice project discipline. Have one general purpose dinking around project but put module-level work in separate projects. If you don’t have one already, have it walk you through setting up a free GitHub account and get into the habit of creating project repos to save your work. You’ll find that invaluable to clawing your way back to the last working stage. You have a choice of models. Currently, Sonnet is the sweet spot for general work. It’s faster and cheaper. Fable is overkill. Opus, occasionally will be helpful for more complex work.
A project is more than a folder. Have Claude “generate a Julia project” and you will stay further away from dependency hell than if you do everything in your global environment.
In the beginning of a project use the /init command. You can scroll through other commands with just the / prefix. Claude is its own help system. “Help me set up a workflow for coding in VSCode in Julia” or “set up sharing helpful functions across projects.”
Periodically ask it to review the project and suggest improvements.
Have it add docstrings to functions.
You will seldom have to resort to Google if you do this.
I asked Claude how to do this. I use VS-Code but not Pluto.
Here is what I asked:
I have been working with Opus 5.5 (medium) to develop the julia package I maintain by giving manual prompts, responding manually, and copying code returned into VS code again manually. I want to start using agents to make the development process more efficient, so that Opus can edit, run, test, adjust itself in a sandboxed environment but still discuss with me key design choices and other decisions and never makes any PRs to my main repo - I will always do that manually. Edits are entirely local. Can you advise how best to set this up? My main tool is VS Code.
Claude gave me very clear guidance which I followed and it’s worked great. Essentially not much more than adding the Claude extension and separating plan and auto work. The best advice it gave was about avoiding git interactions I didn’t want - No PRs, etc. Don’t rely on instructions for this. Give the agent its own clone that has no GitHub credentials. Fetching a public repo still works, but pushing and opening PRs cannot happen. Deny rules are then a second layer. I pull finished branches into my primary clone and push from there manually.
It suggested using WSL for complete sandboxing and Kaimon to give the agent a persistent live Julia session, with tools for running code, running tests with pattern filtering, introspection and debugging. Set-up for these was more involved, so I didn’t adopt these suggestions - too unfamiliar, too advanced! Maybe I’ll get there in future.
