Prompt & Done · LLAMACODE pack

Claude Code for $0: run it on a free model on your own laptop.

Claude Code is the agent: it reads your folder, writes files and runs commands. Ollama lets it talk to an open model running on your own machine instead of a paid API. This is the setup from the reel, the exact prompt, and what we measured on 22 Oct 2026.

1 The commands

Install Ollama (free, MIT), pull the model (6.6 GB), install Claude Code, then launch it through Ollama:

# macOS / Windows: download the app from https://ollama.com/download
# Linux:
curl -fsSL https://ollama.com/install.sh | sh

ollama pull qwen3.5

# macOS / Linux
curl -fsSL https://claude.ai/install.sh | bash

ollama launch claude --model qwen3.5

Prefer to set it up by hand? This is what ollama launch claude does for you (from Ollama's Claude Code page):

export ANTHROPIC_AUTH_TOKEN=ollama
export ANTHROPIC_API_KEY=""
export ANTHROPIC_BASE_URL=http://localhost:11434
claude --model qwen3.5

2 The prompt

Make a folder, start it in that folder, and send:

Make a one-page website for a farm stand, sized for a phone. One file: index.html.

Then open game.html in a browser.

3 What we measured (22 Oct 2026)

Machine: Apple M4 Pro, 24 GB. Ollama 0.40.6, qwen3.5:latest = Qwen 3.5 9B (Q4_K_M, 6.6 GB, Apache 2.0). Claude Code 2.1.308 with a fresh, empty home folder (default tools, no plugins), pointed at Ollama with the env vars above. ollama ps showed 100% GPU and a 64k context.

Job	Free local (Qwen 3.5 9B)	Paid cloud (Claude Sonnet 5.5)
Farm stand website, 3 runs	3 of 3 worked, 195-196 s, $0 API bill	3 of 3 worked, 14-15 s, $0.04-0.05 API-equivalent
Phone games (Breakout / Snake), 8 runs	1 of 8 clearly playable and complete; the rest had a runtime error, a frozen loop or a clipped paddle	Breakout 3 of 3 worked, 28-31 s, $0.09

Cloud costs are what Claude Code reported in our own session (API-equivalent). The local $0 is the API bill only: your hardware and electricity are not free.

4 Tips that matter

Context length. Ollama's docs say to use a local model with tool support and a 64k+ context window. On our 24 GB Mac Ollama allocated 64k by itself (ollama ps shows CONTEXT). If yours shows 4k, start the server with OLLAMA_CONTEXT_LENGTH=64000 ollama serve.
Keep jobs small. One file, one clear goal. A 9B model drifts on big refactors.
Expect retries. Web pages worked every time for us; small games often did not (table above). If a run is broken, run it again: it costs nothing but minutes.
Use a bigger local model if you have the memory (the library lists 27b and 35b sizes for Qwen 3.5). We did not test them.

5 Honest limits

This is Claude Code (Anthropic's agent) driving Qwen 3.5, an open model from Alibaba (Apache 2.0). It is not Claude the model, it is slower, and it makes more mistakes than the paid cloud model. "Free" means no API bill: you still need the hardware (our Mac has 24 GB), and the electricity. We measured one Mac, one model, one kind of job.