VibeKit
Enter App
Free AI Coding Agent
VibeKit's Qwen provider runs Qwen3 Coder on OpenRouter's free pool — $0 per token, $0 platform markup. The only genuinely-free AI coding agent that ships real apps to real URLs. No credit card. No trial period.
Why this works (and why it's actually free)
OpenRouter operates a "free models" pool sponsored by the model providers themselves — Alibaba publishes Qwen3 Coder on OpenRouter at $0/token as a marketing investment for the open-weight Qwen family. VibeKit routes Qwen requests through OpenRouter on a gateway-level key. Our platform markup is 20% over OpenRouter cost. 20% of zero is zero. You get the agent, the container, and the deployment for $0.
How good is Qwen3 Coder, honestly?
Roughly comparable to Claude 3.5 Haiku and GPT-4o on coding benchmarks. For the everyday agent workflow — "fix this bug", "add a route", "tweak this CSS", "scaffold a CRUD page" — it's perfectly capable. Where it lags is large-context reasoning (very long files), complex architectural decisions, and certain niche frameworks. For those, switch to Claude Sonnet or GPT-5 per-app — five seconds in Settings.
The honest comparison
| Provider | Per-call cost | Best for |
|---|---|---|
| Qwen3 (free) | $0 | Everyday agent loops, prototypes, tweaks |
| DeepSeek v3.1 | ~$0.04–0.10 | High-volume / cost-sensitive serious work |
| Claude Sonnet | ~$0.10–0.30 | Best reasoning, large refactors, complex bugs |
| GPT-5 | ~$0.10–0.30 | Best for OpenAI-flavored ecosystems |
Per-call costs are typical agent tasks ("fix a 500", "add a route") at platform-fallback pricing (OpenRouter cost + 20%). With BYOK on Anthropic / OpenAI you pay the provider directly — VibeKit's markup goes to zero too.
Switching to Qwen
- Open VibeKit on iPhone or web → Settings → AI Provider
- Pick Qwen from the four options (Anthropic / OpenAI / DeepSeek / Qwen)
- Your next agent message routes to
openrouter/qwen/qwen3-coder:free - That's it. No key, no payment, no setup.
What's still included
- Real hosting. 1 app, 256 MB RAM, 1.5 GB workspace, on AWS Fargate at
yourname.vibekit.bot. - Real GitHub. Auto-created repo, auto-redeploy on push, full Git history.
- Real SSL. Let's Encrypt cert from second one.
- The iOS app + Telegram bot — same UX whether you're on Qwen, Claude, or anything else.
When to upgrade
If you find yourself bumping into Qwen's limits on a specific task ("the agent isn't getting this architectural change right"), switch that one app to Claude Sonnet for the day, fix the thing, switch back. With BYOK on a Claude Pro / Max subscription you can even do this for free. Different apps can use different providers — granular, not all-or-nothing.