Published: 2026-10-21
Summary

Free Claude Code, an independent proxy: route Claude Code to Ollama or OpenRouter free models, with fallback and per-tier overrides

Chapters / key moments (click to jump — plays here on the page)

FCC is an independent open-source project, not affiliated with Anthropic, that listens on your machine and forwards Claude Code's requests to a model you choose, including local Ollama models and free OpenRouter models. Fahd Mirza installs it, sets a default model plus a fallback, and stops Ollama mid-task to see whether work continues. It is a useful pattern, and Anthropic's documentation is clear that this route is unsupported.

Source video

"Free Claude Code: Can You Really Run AI Coding Agents for Free?" by Fahd Mirza — Watch on YouTube →

What the video shows

Install. A one-line installer from the project's repository (linked in his description). It asks whether to also set up Codex and which providers to configure; he declines the extras. Then he starts the proxy server and leaves that terminal open; a dashboard opens on his machine.
Providers. The dashboard detected an OpenRouter API key already set in his environment and listed 396 free models, and found his running Ollama. He points out that picking up an existing key could be a security concern. Any other OpenAI-style endpoint on your network can be added as a custom provider.
Model config. One default model handles every request unless you override a tier. Claude Code asks for Fable, Opus, Sonnet and Haiku tiers; an override sends one tier to a different model, and "none" means that tier uses the default. Press Apply at the bottom right; if the Ollama model doesn't appear in the list, type its name.
Launch. In another terminal he runs the project's clone launcher and Claude Code starts as usual. It still labels the model "Opus 5.5" although no Anthropic key is set and nothing goes to Anthropic: the request is caught by the proxy and answered by his local model.
Request routing and fallback. Per the project's readme (as he reads it), small housekeeping requests such as generating a title are answered locally without any model, and everything else goes to the selected model. If that model fails after retries, the proxy moves to the next configured model in the same turn. He tests it by stopping Ollama while a task runs; the task carries on through his OpenRouter free fallback.
Extras. The dashboard also takes Telegram and Discord tokens and has integration entries for editors and other harnesses.

Checked against Anthropic's docs

The Claude Code gateway page says ANTHROPIC_BASE_URL is the variable that points Claude Code at a gateway. It also says Anthropic "doesn't endorse, maintain, or audit third-party gateway products, and doesn't support routing Claude Code to non-Claude models through any gateway." Treat an arrangement like this as unsupported.
Per the same page, setting only the base URL without a gateway credential doesn't replace a claude.ai subscription: a saved login stays the active credential, so its usage limits and billing still apply.
The video's "over a billion free tokens a month" figure comes from the project's readme. He says so himself and tells viewers to expect rate limits, throttling, and weaker results on complex coding tasks.

Gotchas & Caveats

Free tiers send your prompts to the provider. Check each provider's terms before using private code.
Claude Code sends features that non-Claude models may not support; expect odd behaviour that is not a bug in either tool.
Install scripts run with your user's permissions. Read the repository before running a one-liner, and note that a proxy sees every prompt and file Claude Code sends.
We did not review the project's code, so the routing and fallback behaviour is as shown on screen.

Key Takeaways

The pattern is a local proxy with a default model, optional per-tier overrides and an ordered fallback list, so one failing provider doesn't stop a task.
It is unsupported by Anthropic, and a missing fallback key or a stopped local model are the first things to test.

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

Install the proxy
A one-line installer asks which providers and agents to set up; the speaker declined Codex and every provider he did not use. It then starts a local proxy server that must stay open in its own terminal and opens a dashboard on the local machine.
Pick providers in the dashboard
It auto-detected an OpenRouter key already present on his system (listing 396 free models) and a running Ollama (Qwen 3.8). Custom endpoints on your network can be added.
Set the default model and per-tier overrides
In model config the default model handles everything. Claude's tiers (Fable, Opus, Sonnet, Haiku) can each be overridden to a different model; an override set to none uses the default. Click apply, and type the model name by hand if it does not appear.
Launch the client in a second terminal
The speaker starts the proxied agent in another terminal (he calls it the FCC clone). The agent still displays Opus 5.5 but its requests go to the local model, with no Anthropic key set.
Add fallback models
Choose fallback models in the dashboard. If the main model fails after retries, the proxy moves to the next in the same turn. He tested it by stopping Ollama mid-task; the run continued on the OpenRouter free model.

Gotchas

Independent open-source project, not affiliated with Anthropic or any other company. The README's claim of over a billion free tokens a month is unverified, and free tiers are set by each provider and can change.
Expect rate limits and throttling, and slower, weaker models on complex reasoning tasks. Check each provider's terms for how prompts are handled before sending private code.
The speaker flags a possible security issue: it picked up an OpenRouter key from his environment without being asked.
Small housekeeping requests such as generating a title are answered locally by the proxy without calling any model, per the README as read in the video.
Messaging tokens for Telegram and Discord can also be entered in the dashboard. Exact command names are not shown on screen in the transcript; use the project's README.
The video's title mentions what Anthropic supports, and this proxy is not an Anthropic-supported way to run Claude Code.