Free Claude Code, an MIT-licensed local proxy that lets Anthropic's Claude Code run against roughly 50 other model providers, has passed 49,000 GitHub stars and 8,000 forks. It works by accepting Anthropic's own wire protocol from the client and translating it to whatever upstream you configure, so Claude Code never knows the model changed. The popularity is the story: users are voting, at scale, that the harness is the thing they wanted.
Key facts
- The repository sits at 49,044 stars and 8,005 forks, under the MIT license.
- It supports roughly 50 providers and 9 coding agents, per its own README.
- Routing is per tier:
MODEL_FABLE,MODEL_OPUS,MODEL_SONNET, andMODEL_HAIKUcan each point somewhere different. - Anthropic's own documentation confirms the underlying pressure: usage limits vary by plan and model, and web, Desktop, and Claude Code all draw from the same pool.
Ask a developer what they pay Anthropic for and most will say "Claude." Ask them what they would miss if the model changed but the tool did not, and the answer gets more complicated. Forty-nine thousand people have now installed the experiment.
Here is how it works. Claude Code speaks Anthropic's Messages API. The proxy runs on your own machine, accepts that traffic as if it were Anthropic, and forwards each request to a provider you choose -- an NVIDIA endpoint, OpenRouter, Gemini, DeepSeek, Groq, Cerebras, a local Ollama or llama.cpp server, and so on. The architecture document is explicit that this is not a dumb pipe: the Anthropic-facing layer owns content blocks, tool-call structure, thinking blocks, and server-sent-event streaming, while separate provider adapters handle the upstream conversion. That protocol fidelity is the entire trick. Claude Code is a demanding client; if the tool-call shape or the streaming envelope is wrong, it breaks visibly.
The detail that makes this more than a cost hack is tiered routing. Claude Code does not use one model. Internally it asks for different capability tiers depending on the job -- heavy reasoning for planning, something cheap for a quick classification. The proxy exposes each tier as a separate setting, so you can send the expensive tier to Kimi or GLM, the middle tier to a fast Gemini model, and the trivial tier to something tiny on Groq, with reasoning effort configurable per tier. The result is a harness with a heterogeneous fleet behind it, which is closer to how a cost-conscious engineering team would design this from scratch. Ground Truth's explainer on model routing and cascades covers the general pattern.
Think of it as a universal power adapter. The appliance is the same; the wall socket changed; a small box in between makes the mismatch invisible. That framing also explains why the project keeps growing beyond its original name -- it now fronts Codex, OpenCode, Cline, and several other clients, which makes it less a Claude workaround than a general adapter layer for coding agents.
The demand side is not mysterious. Anthropic's own support documentation states that usage limits vary by plan, model, feature, and effort level, and that activity on Claude's web interface, Claude Code, and Claude Desktop all counts against the same allowance. Its plan page describes Max 5x and Max 20x as multiples of Pro capacity per session. Users have been reverse-engineering what that means in practice, with mixed feelings. In issue 76133, a Max 20x subscriber describes a long build hitting a limit with no advance warning and asks for a visible remaining budget and per-subagent consumption. In issue 87419, another Max 20x user reports weekly meters depleting far faster after the August 17 reset, and an OAuth token briefly carrying the wrong rate-limit tier. Anthropic did recently remove the five-hour usage limit in a related move on the other side of the market.
Why it matters: this is the same thesis OpenAI just endorsed from the opposite direction when it open-sourced the Codex harness. One company gave away its loop deliberately; Anthropic's users took theirs by force. Both point at the same conclusion -- the harness is the product, and the model is becoming a configurable input. See agent harnesses and scaffolding for why the loop shapes so much of what an agent can actually do.
Two honest caveats. First, the name oversells it: nothing here is Anthropic's model for free, and the strongest defense of Anthropic's position is simply that these are intentional plan-based allowances with paid headroom available, not a defect. Second, running a local proxy that holds credentials for fifty providers and terminates your coding agent's traffic is a real trust decision. The license is permissive and the code is readable, which is the mitigation, but it is not the same as no risk -- and as the essay on inference-engine attack surface argues, every extra parsing layer between a model and a machine is another place for a bug to live.
Originally published on Ground Truth, where every claim is checked against the primary source.
