Never
Claim unlimited free Claude
Premium models cost real money per token. You earn access in credits, capped and visible. Anyone promising infinite frontier models for free is either burning investor cash or about to change the deal.
✓ Email confirmed - thanks, you're set.
Free AI in your editor, funded by the companies that want to hire you.
DeepSeek and Qwen run free, every day, no strings. Claude, GLM and Kimi run on credits you earn by watching one sponsor spot. Every credit that comes in and every token that goes out is written to a ledger you can read, and the balance can never go below zero.
This is what it feels like inside your editor.
A mock of the real thing, built by us, on this page. The file tree, tabs and menus are for show. The panel on the right actually runs: pick a model, chat with it, and watch what happens when you reach for a credit-gated one without enough balance.
4devs assistant
40 cr
deepseek-flash
I can see
app.js open. Ask me to explain it, find a bug, write a test,
or suggest a refactor, or just open a different file first.
Not enough credits for this model yet.
Scripted for this preview. Nothing here calls a real model or leaves your browser.
Four things this product will never do.
Most "free AI" tools in this space are free until the seed money runs out. We would rather publish the constraints first and be judged against them.
Never
Relabel a cheap model as Claude
The model name in the sidebar is the model that answered. If you ask for Sonnet and get Sonnet, the ledger says Sonnet. Silent downgrades are the oldest trick in this market.
Never
Resell someone's subscription
We are not a grey-market reseller of Claude Pro or ChatGPT Plus seats. We buy API access at list price and wrap it in our own product, with our own terms.
Never
Read your source code
Ad matching happens on your machine, from your manifest file: package.json, requirements.txt, go.mod. What leaves is a short list of tags, not your repository.
Two tiers. One of them is free for everyone.
The free tier runs on models cheap enough that ad revenue covers them several times over, so it stays on permanently, with only a daily cap to stop scripted abuse. GLM-5 joins that tier: strong enough for real agentic work, and just as cheap to run as DeepSeek and Qwen. Claude and Kimi cost real money per token, so those two run on credits instead.
| Model | Tier | What it's good at | What it costs you |
|---|---|---|---|
| deepseek-flashDefault model | Free | Everyday completion, refactors, tests, boilerplate, "explain this function". | Nothing. Generous daily cap, resets at midnight EET. |
| qwen-coderFree alternative | Free | Long-context reading, multi-file edits, non-English comments and docs. | Nothing. Shares the same daily cap. |
| glm-5Free alternative | Free | Agentic multi-step work, 200k context, repo-wide refactors. Open-weight and close to frontier on coding benchmarks. | Nothing. Shares the same daily cap. |
| minimax-m3Free alternative | Free | 1M-token context and native multimodal input, useful when a question involves a screenshot or a long design doc alongside the code. | Nothing. Shares the same daily cap. |
| step-3.5-flashFree alternative | Free | Fast everyday completions and math-heavy snippets, priced low enough that it barely moves the ad-revenue math at all. | Nothing. Shares the same daily cap. |
| claude-sonnetPremium | Credits | Hard debugging, architecture, gnarly legacy code, anything where the free models keep guessing. | Credits, priced at real token cost plus a thin margin. Balance shown before you send. |
| kimi-k2Premium | Credits | Very long files and repo-wide context at a fraction of frontier pricing. | Credits, at roughly half the rate of Sonnet. |
deepseek-flash
FreeWhat it's good at
Everyday completions & refactors.
What it costs you
Free. Daily cap, resets midnight.
qwen-coder
FreeWhat it's good at
Long-context, multi-file edits.
What it costs you
Free. Same daily cap.
glm-5
FreeWhat it's good at
Agentic work, 200k context.
What it costs you
Free. Same daily cap.
minimax-m3
FreeWhat it's good at
1M context, multimodal input.
What it costs you
Free. Same daily cap.
step-3.5-flash
FreeWhat it's good at
Fast completions, math snippets.
What it costs you
Free. Same daily cap.
claude-sonnet
CreditsWhat it's good at
Hard debugging, architecture.
What it costs you
Credits, real cost + thin margin.
kimi-k2
CreditsWhat it's good at
Long files, repo-wide context.
What it costs you
Credits, ~half Sonnet's rate.
1 spot
One sponsor slot, one balance top-up
Shown while the model is already thinking, or when the editor has been idle. Never mid-keystroke.
0
Lowest possible balance
Your balance is checked before the request is sent, not after the tokens are spent.
~10%
Margin we take from the pool
The rest of the ad revenue goes straight back out as premium model access.
3
Editors supported at launch
VS Code, Cursor and Windsurf share one extension API. JetBrains is a separate codebase and comes later.
Reach developers where they're already working
4devs is a free AI coding assistant inside VS Code. Sponsors show up in two places while a developer is actually using the tool, never as a popup, never dismissible for a free-tier user. Below is exactly what each placement looks like inside the real editor. Click any ad to try building your own, no account needed.
1
Passive text
Shown inline next to the "thinking" indicator while a reply streams in. Lowest friction, lowest CPM.
2
Passive banner
A fixed, non-dismissible card in the sidebar for free-tier users. More real estate, still zero interaction required.
Why is this retry loop never resetting its counter?
Thinking…
function retry(fn, max) {
let attempts = 0;
while (attempts < max) {
// attempts never resets between calls
attempts++;
}
}
The awkward ones first.
Q1 What's the catch?
The free tier runs on models that cost a fraction of a cent per request, so ad revenue covers them comfortably. That part is genuinely free and stays free. GLM-5 costs more than that but nowhere near frontier prices, so one sponsor spot covers a whole session of it.
The catch, if you want to call it that, is the premium tier: Claude and Kimi cost real money per token, so you pay for them with attention instead of euros. If you never watch a sponsor spot, you never get premium access, and you keep the free models anyway.
Q2 Do you read my code?
No. Ad matching reads your dependency manifest (package.json, requirements.txt, go.mod) on your machine, and turns it into a short list of tags like react, postgres, aws. Those tags and your country are what get sent, anonymously.
Your prompts do go to the model provider you selected, because that's how inference works. That's true of every AI coding assistant, and any tool that tells you otherwise is being careless with the word "private".
Q3 Is the Claude here the real Claude?
Yes. Anthropic's API, at Anthropic's prices, which is exactly why it is credit-gated rather than free. The model that answered is named in the ledger entry for that request.
What we can't do, and won't pretend to do, is resell a Claude Pro subscription at a discount. Nobody can. If someone offers you that, ask what they're actually routing your requests to.
Q4 Why not a Chrome extension?
Because the audience is the product. A verified developer in a specific city working in a specific stack is worth several times a generic browser user to the recruiters and companies who fund this. Opening it up to everyone with a browser dilutes exactly the thing that makes the ad slots worth buying.
Q5 PyCharm? IntelliJ? Neovim?
Not at launch. VS Code, Cursor and Windsurf share one extension API, so supporting all three is nearly free. JetBrains is a separate codebase, a separate review process and a separate maintenance burden. It gets revisited once v1 has proven revenue, not before.
If you're on JetBrains, join the waitlist and pick your editor anyway. That count is how the decision gets made.
Q6 What happens when my credits run out mid-task?
The request is refused before it's sent, and you're offered the free model instead. The balance floor is zero, checked server-side. No overdraft, no "you owe us", no card on file to charge.
Q7 Why Greek and Balkan developers specifically?
Two reasons, one idealistic and one commercial. Frontier model subscriptions cost the same in Athens as in San Francisco while salaries do not, so the price of good tooling bites harder here. And the regional ad market is underserved enough that a company hiring in Thessaloniki genuinely cannot buy this audience anywhere else.
The site launches in English and Greek. Serbian and others follow based on where waitlist signups actually come from. Building translations for a market that hasn't shown up yet is how small projects run out of time.
Q8 Isn't this just Idlen, Kickbacks or AdVibing?
Those tools pay you cash for ad impressions while you use someone else's AI. This is the other way round: the ad revenue is spent inside the product, on model access, which is why the free tier can exist at all rather than being a discount on a subscription you're still buying.
The other difference is where the ads come from. Their inventory is global and programmatic. Ours is meant to be sold directly, locally, at rates that only make sense because the audience is specific.
Want to add something to this list? questions@4devs.eu.
The extension is the easy part. Proving the audience is the work.
That's what the waitlist is: not a launch countdown, but the evidence that decides whether any of the rest gets built. Four fields, thirty seconds.