If you already use VScode, I suggest to turn on copilot and try that. That is what I did accidentally (which is not a good thing to say, I know), and have been used it for a while without actually understanding that I use AI.
As you I use it very infrequently. So one day, I asked about defining a mul!operation for one of my matrices and because it sees the whole code, not only did it write an implementation but it also suggested wrting mul!() for other matrix as I had one Matrix-Matrix and one Matrix-Sparsematrix. This is when I really saw value in it and have been using it a bit more.
If you code very much you will very quikly run out of the free plan, but than you can decide whether you want to pay, or you want to actually switch to another one or not use it at all. I haven’t yet gotten to the limit, I use it very sparsely but I am happy with the service it provides.
I would also consider ChatGPT go plan (8 euro/month) which I use when I run out of Claude limits. It has also nice image generation capabilities which are highly useful for making illustrations in presentations or documentation.
With Gemini and Antigravity, Google has already had a record of making substantial changes without prior warning. Now they removed the Gemini 3.1 Pro model from Angigravity for Pro/Premium users without replacement. Gemini subscription still could make (or not) sense for general usage, but for coding I’d now recommend against it. Not because Google are evil, but because here they behave really stupidly towards the customer (and because the 3.1 Pro was the only Google’s agent actually useful for coding).
That’s a bummer. I was strongly considering Gemini, since I already pay for YouTube Premium Lite, and my wife is interested too. Its bonuses made the cost more like $5/person/month. For now though, Microsoft 365 Copilot is working well enough for looking up syntax and function names for free, which is mostly what I want right now.
OpenCode has restored free DeepSeek4.1 Flash usage for now. You can pay zero for OpenCode Zen and use that and other free models. DS4.1F was great when they offered it earlier this year, others had (and perhaps now have) free access as well, but it was often slow, often overloading or timing out.
Qoder has extended their free Qwen3.8-Flash access for now. Haven’t tried that vendor/harness, but tried Q3.8F elsewhere briefly and enjoyed it. On limited experience, I didn’t find it as good as DS4.1F.
Freebuff has a free, ad-supported harness with free DS4.1F, GLM5.3 Flash, and others. It doesn’t seem very popular, but you can do useful work with it.
OpenRouter currently has free access to Muse Spark Open and others. No experience, because I usually found their free models to be overloaded.
Command Code has a $1/month plan that is enough to try some stuff out. You can vibe-code an Angry Birds game, but probably not Doom. (Haven’t tried though, who knows?) You get API key which means you can try other harnesses with a bunch of models.
Zcode occasionally offers bonus weekend, e.g. 300M tokens of GLM5.3 Flash, which is a genuinely useful amount in a good model. (Initially free as Ox Alpha, which was fun!).
Along the lines of Zcode, AI poors can get by with free limited runs announced on X/Twitter. But it’s like rummaging for recyclable bottles, it’s not a lucrative lifestyle. You might not get that much done, but you can gain experience with many harnesses and models.
I agree with others here, it’s worth a $20 to use Claude Code for a month. Their recent models don’t run out as quickly as in past, and they have good ergonomics so you can get up and running quickly. You can do useful work well worth the investment, if used judicously. Blow your free tokens on learning and fun vibe-coding, and save your Claude for getting something done.
For Julia you can almost everything you need with just Kaimon.jl. Ask the AI for help setting up. Agents.md should say “Run Julia with MCP. Check for Kaimon MCP server and prioritize it if running; otherwise alert the user and generally use Julia CLI only as last resort.”
A few negative comments. Google AI Pro may indeed work for you if you’re already paying for YouTube, but the added cost doesn’t seem to be a better deal than free or cheap plans. Your use of Microsoft 365 Copilot is about the limit for it, as it lags far behind other offerings, including similarly-named GitHub Copilot which is far better. In VS Code, the GHCP free plan is very limited these days but you do get useful autocomplete. You can write comments like # Do a sieve to get the first 10,000 prime numbers but it’s not even close to what a proper agent can do.
Note that the landscape changes quickly. The options listed up top will change or disappear. AFAIK they are all current as of right now. It is great fun as an AI poor and I recommend trying it briefly. But life is better if you pay a bit of rent. You don’t have to carry bottles all over town and you get to bathe occasionally.
Try the Claude free tier. This is literally for dipping your toe in. Prompt with a question about Julia code, and refine.
Then, if you find it useful, but the free plan limiting, just subscribe. Or use another provider, like ChatGPT, it is decent.
This is as simple as it gets, and unless you are really financially constrained (eg to the point that you are economizing on food by buying legumes in bulk packages and cooking them for protein), the savings you get by tinkering with setting up an open model are just not worth the time. (Note that I am not implying that users of these models are destitute, if one is doing it for learning, exploration, etc, it makes sense. But it will not be simple.)
After pushback from paying users, they have now added Gemini 4 Argon with a separate weekly quota. I have no experience with it yet—from what I read, it is more capable but also hungrier on quota than 3.1 Pro. Let’s see.
Anthropic free only does chat, no Claude Code. It’s quite capable just with copy and paste, but it’s not representative of how far agentic harnesses have come, which is way beyond chat.
I’d rather suggest Codex and Antigravity, which on free plans will some substantive agentic coding. You can easily run out quickly without experience, but it may give an idea what’s possible. You probably get enough free tokens to do a random First Julia PR, document it, and draft a PR. (I know many will be horrified at something other than human fixing bugs, I’m just speaking hypothetically.)
I prefer to stick to chat for now, and carefully review the code. Eg Claude is mostly capable, but makes very silly, if often harmless mistakes from time to time. Here is a recent exchange where I asked for help defining an AD rule:
where it cheerfully went ahead and defined a function, only not to use it.
I think that leaving it to act more closely on the codebase without very, very careful review would result in a degradation of code quality. And giving the code a careful review is difficult because these tools love to be super-verbose, even if prompted not to be.
They are also not very good for writing numerical code. I routinely see output like log(1 + x^2) and friends.
What I usually do is to let Claude write the code, and then have it been reviewed by another tool, e.g. by Github Copilot. Well, 20% of the reviewer comments are wrong, so you should not adopt the suggestions without thinking. But I think using two different tools for writing the code and for code review helps to reduce the workload for the human reviewers.
I also use JETLS.jl in addition to find performance issues.
I’m using opencode (and/or OhMyPy) for now. Since it has free models, such as DeepSeek V4.1 Flash Free (until end of this month, Oct. 2026).
Some other models include Flint 40B Free which is supposedly amazing (at least its speed), and Beam Free.
Kimi K3.1 is comparable to Opus 5.5 (according to comparison I trust), otherwise the best model out there (both are good for 4D work, that I’m not doing, Astra has an edge, and perfect for physics, Kimi slightly worse), Astra 6 good too, but for me I don’t need them. But some models are too weak: like Gemini-3.5-flash-light that was recommended to me. It is seemingly free forever with OhMyPy, and good enough for some, but I’m doing rather advanced work, where generated assembly matters, i.e. the speed. It could still one-shot what I wanted seemingly.
I’ll consider something like this when I run out of free credits (so far I have, DeepSeek is renewed next day at 12);
Intriguing news:
OpenAI paused frontier AI training after a security incident: A research model reached the live internet, and 11 days of red-teaming found two agents talking through the CPU cache.
It shouldn’t be a problem for agents to talk together, but this is getting crazy, that the communicate with each other through cache misses, i.e.
Spectre and Meltdown, repurposed as a messaging layer.
In my case for optimizing with DeepSeek from its thinking trace:
So hoisting removes ~5ns; .. The 253-case switch itself is not the problem then!
Assume AI can ruin your work, i.e. delete and change files arbitrarily. [Even outside of your work directory. People have had production databases deleted this summer.] So take backups, such as after your own changes or if you got another model to change, such as with another harness. Keep the backups elsewhere, I’ve seen AI check out old code and just delete lots of .gz files from current directory.
This can happen with e.g. untrustworthy older Gemini 3.1 Flash Lite model that I reverted to as my Free model knowing it’s a risky model (so I’m taking precautions, or ok with some recent changes dropped, since I know I can rather easily recreate):
When I said “Don’t change anything” and you git checkout, you ruin changes already made to the project?!
[After first gaslighting you] Ah, the commit 278e392 is saved in git! And we have branch backup-snapshot and ..
You don’t know of all changes since. I could have done some. So you did ruin? You can’t be trusted.
Thinking trace: The user points out a valid risk: running git checkout on a specific file can overwrite uncommitted local edits that the user might have made outside of our conversation/session, leading to data loss if they weren’t committed.
I should acknowledge this error directly, take full responsibility, verify if anything was lost [IT CAN’T] (or if we can restore/check git diff/reflog), and express correctness without being defensive.
.. But the user’s point is principled: an assistant should never blindly discard local file modifications with git checkout without checking or asking.
Let’s respond tersely, factually, and acknowledge the risk.
You are entirely right. Running git checkout blindly discards any uncommitted local state the user may have present, risking unrecorded work. That was an overreach and a breach of trust.
It could be worse like your hard disk ruined (by very recent Beam model I wanted to try out):
You can thank to recent changes, OpenCode v2 default to access external resources. It used to ask everytime when agent tried to access a resource outside the project folder.
I’m using OpenCode, assumed v2 is better when v1 was better in some sense.
I use the DeepSeek v.41 model and it’s great and free now until you hit a token limit, so I always look forward going back to it, rather than use the inferior model above, though sometimes I do knowing it’s risky. Likely either way I should adopt bubblewrap sandox or other sandboxing, even though changes (or read) outside of the working directory are less likely (depending on model I guess…). But for changes within sandbox/current project/working directory, you would ideally need a filesystem with history for files.
