Free Claude Code Is Trending on GitHub, and the Interesting Part Is the Routing
An independent project aggregating 50 providers and nine coding agents behind one catalogue is topping GitHub's daily trending list. Its failover design is the part worth studying.
Free Claude Code — FCC — sat at the top of GitHub's daily trending list this morning. It is an independent open-source project, not affiliated with or endorsed by Anthropic, and it does something structurally interesting regardless of what you think of the premise.
What it aggregates
The project puts 50 providers it describes as terms-of-service friendly behind a single searchable interface, covering free, paid, subscription and local models. It claims 1.3 billion free tokens a month across those providers, and says it removes integrations if a provider stops permitting them.
On top of that catalogue it runs nine coding agents: Claude Code, Codex, Pi, OpenCode, Cline, Hermes, DeepSeek Harness, Grok Build and Muse Code.
Free-tier availability and limits are set by each provider and can change, which the project states plainly.
The failover design
The feature worth extracting is what happens when a provider goes down. After retries are exhausted, FCC automatically moves to the next configured model without making you restart the turn — and it does this across every client.
That is a harder problem than it sounds. Preserving conversation state and tool context across a mid-turn provider switch is exactly where most multi-provider abstractions break. Whether FCC gets it fully right is a separate question, but naming it as the design goal is the right instinct.
Token economics
FCC claims up to 90% fewer terminal-output tokens. An optional component called RTK filters common command output, and five other optimisations handle quota probes, command-prefix detection, titles, suggestions and filepaths without calling a provider at all.
Terminal output is genuinely one of the largest and least useful token sinks in agentic coding. Handling it locally rather than paying a model to read it is sensible.
Surfaces and capabilities
It runs through native launchers, VS Code, the Codex App, JetBrains, Discord and Telegram. Voice input works through local Whisper or NVIDIA NIM transcription.
The project says agent capabilities stay intact: streaming, tool use, native interleaved thinking preserved, image input, and independent routing of Fable, Opus, Sonnet and Haiku to compatible models.
A note on the install
Installation is a piped shell script on macOS and Linux, and the PowerShell equivalent on Windows. The project links both installers for review before running them, which is the right thing to offer. Read them.
Why it is trending
There is obvious appeal in "free tokens", and that explains the star count. The more durable idea is the one underneath: a provider-agnostic routing layer with real failover, sitting between developers and a market where models, prices and availability change monthly. That abstraction is going to exist in some form. This is one attempt at it.