Vibe-coding for free

Ready-made: code with AI without paying for subscriptions. A full agent inside the editor reads your repo, edits files and runs commands — on free or local models. Built entirely from open-source cards.

What you get

• An environment where the AI agent sees the whole project, edits files and runs the terminal.
• Free (or local, private) model access — no monthly subscription.
• Current API docs in context, so the AI does not invent stale methods.
• A real browser check and proper Git-branch workflow.

What you need

Tools from the base: VS Code (editor), Cline (agent), OpenRouter or Ollama (model), the Context7 MCP (fresh docs), the Playwright MCP (browser check), Git (versioning). Free with an OpenRouter key or fully local with Ollama (a GPU 8GB+ / Apple Silicon 16GB+ helps for local).

Step by step

1. Editor — install VS Code, open the project folder. OK when the tree and terminal show up.
2. Agent — install Cline (open-source): reads the repo, edits files, runs commands on confirmation. OK when its sidebar icon and chat appear.
3. Model — OpenRouter (free models, one key) or Ollama (ollama run qwen2.5-coder). Put the key/endpoint into Cline settings. OK when a question gets an answer with no auth errors.
4. Fresh docs — the Context7 MCP: add use context7 to a request so the agent uses real APIs. OK when the code uses current methods, not invented ones.
5. Browser check — the Playwright MCP: ask the agent to open your local URL and walk the flow. OK when it actually clicks through and reports pass/fail.
6. Versioning — Git + branches: git checkout -b feature/<task> before work, run the "Code review by checklist" prompt before committing. OK when git status shows the new branch.

Free vs fast (paid)

Free/private: Ollama or OpenRouter free models, speed depends on your GPU. Faster: a top model via OpenRouter for tokens — quicker and smarter on hard tasks. For an agent team on top, see "Laskoff OS".

Common problems

Local model is slow → use a smaller (7B) or quantized model, or switch to an OpenRouter free model. Invented API → always add use context7. Rate limits → rotate free models, split heavy tasks. Wrong edits → work on a branch, git restore to undo. Ollama ignores the GPU → check drivers, run a smaller model.

Time & money

Setup 20–40 min, then you just work. Cost: $0 on OpenRouter free models or fully local Ollama; pay per token for a top model only if you hit a quality/speed ceiling. Honest note: free/local models trail top cloud ones on complex tasks.