Run Coding Agents for Free — Safely (free-claude-code + Ollama)

The Claude Code interface is one of the best coding-agent UIs ever shipped. The subscription is what stops most students. free-claude-code (41k+ stars, MIT) removes that barrier: it is an open-source local proxy that lets the Claude Code, Codex, and Pi interfaces talk to 25+ different model providers — free cloud tiers, or a model running entirely on your own machine.

One thing before anything else, and it is the most important lesson in this guide: this is a proxy that sits between you, your prompts, and your API keys. You read a tool like that before you run it. We will do exactly that.

Two honest scopes up front. First, this is not “free Claude” — unless you plug in your own Anthropic key, it routes the Claude Code interface to other models (free tiers or local ones). Second, “free” means the proxy is free and you use a provider’s free tier or your own laptop’s compute.

Step 0 — Read before you run (the 60-second check)

Every tutorial tells you to pipe a script from the internet into your shell. Almost none tell you to look at it first. Here is the fast, repeatable check — download it instead of piping it:

curl -fsSL https://raw.githubusercontent.com/Alishahryar1/free-claude-code/main/scripts/install.sh -o install.sh
head -40 install.sh
grep -nE "sudo|keychain|security |eval|rm -rf" install.sh

You are looking for three things: where it downloads from (this one: the GitHub repo zip, plus the official installers for uv, Claude Code, Codex and Pi — all first-party URLs), whether it wants root (it does not use sudo), and whether it touches credentials (no keychain access). Sixty seconds, and now you are running something you have actually seen. When we did this check ourselves, the script came up clean. Then, and only then:

sh install.sh

Step 1 — Start the proxy

fcc-server

Watch the startup log — it tells you the important fact:

INFO:  Admin UI: http://127.0.0.1:8082/admin (local-only)

Local-only. The proxy and its Admin UI run on your machine, on localhost. Your prompts flow through this local process to whichever provider you pick — not through some third-party server in between. Your provider keys are stored on your machine (in ~/.fcc/.env). That is the architecture you want from a tool like this — and it is why reading the code first was worth it.

Step 2 — Pick your free brain

Open the Admin UI from that log line. Two good free paths:

Path A — a free cloud tier. NVIDIA NIM, Google AI Studio (Gemini), and OpenRouter’s free models all work: create a key on the provider’s site, paste it into the matching field in the Admin UI (for example NVIDIA_NIM_API_KEY), pick a model from the MODEL dropdown, click Validate, then Apply. Strong models, zero cost, but your prompts do go to that provider under their free-tier terms.

Path B — fully local with Ollama (nothing you type ever leaves your machine). This is the path from our video:

brew install ollama
ollama serve          # leave this running
ollama pull qwen3:8b  # ~5 GB, a solid local coding model

In the Admin UI set MODEL to ollama/qwen3:8b, Validate, Apply. Done — the “provider” is your own laptop.

Step 3 — Launch it

fcc-claude

The familiar Claude Code interface opens. Now the honesty step: type /model and pick the gateway entry — you will see it listed, for example ollama/qwen3:8b — From gateway. The banner switches to show exactly what you are running. Give it a task and watch a local model write real code in a first-class agent UI, with no subscription anywhere.

What to expect from a local model (the honest part)

An 8-billion-parameter model on your laptop is not Claude. In our testing it is genuinely good at focused tasks — write a function, explain an error, draft a test — and it answers in seconds. It gets shaky on big multi-file refactors and long agentic tool chains. So use the free/local path for learning, drilling, and small tasks, and you will lose nothing. When a task really needs a frontier model, you will know — and you can point the same interface at a free cloud tier or a paid key without changing tools.

Key hygiene (do these three things)

1. Use dedicated keys. The key you paste into any proxy should be created for that proxy — never your main production key. Every provider lets you make multiple keys.
2. Prefer free-tier keys with hard limits. A key that can’t spend more than $0 is a key you never have to worry about.
3. Rotate if unsure. Deleted a tool? Rotate the key you gave it. Thirty seconds, total peace of mind.

The habit that matters

“Free” tools that touch your prompts and keys are everywhere now, and most of them are wonderful. The skill that keeps you safe is not avoiding them — it is the sixty-second read before you run, local-first defaults when privacy matters, and keys that are scoped to be worthless if they leak. Learn that habit on a tool like this, and you carry it into every production system you will ever touch.