The free-claude-code proxy1 intercepts Claude Code’s Anthropic API calls and reroutes them to NVIDIA NIM, OpenRouter, DeepSeek, LM Studio, llama.cpp, or Ollama. The project first surfaced on Hacker News in February 2026 as ‘claude-code-free’ with 3 points and zero comments. By late April 2026, the renamed repository was reporting 16.9k stars and 2.4k forks. Anthropic’s subscription enforcement fight over OpenClaw provided the audience the February post didn’t reach.
The April OpenClaw Crackdown and Reversal
On April 3, 2026, Anthropic notified users2 that Claude Pro and Max subscriptions would no longer power third-party CLI harnesses effective the following day, citing “outsized strain on our systems.” Over 135,000 OpenClaw instances3 were active at that point. Affected users got a one-time credit equal to their monthly plan cost, redeemable by April 17.
The enforcement didn’t stay clean. On April 10, Anthropic temporarily suspended OpenClaw creator Peter Steinberger4 for “suspicious” activity. The ban lasted hours before public attention reversed it. By April 21, Anthropic staff told the OpenClaw project that OpenClaw-style CLI usage is allowed again5, with API keys described as “the clearest and most predictable production path.”
The reversal is narrower than the headline suggests. The policy fight was about subscription-powered harnesses consuming Claude capacity at flat monthly pricing. API-key-based usage that bills per token was the preferred outcome all along. free-claude-code operates entirely on the API-key side: it doesn’t proxy subscriptions, it replaces the backend entirely.
How free-claude-code Intercepts Claude Code
The proxy sits between Claude Code and any provider backend, presenting an Anthropic-compatible API surface inbound and translating to each provider’s format outbound. Claude Code’s configuration points at the proxy’s address rather than Anthropic’s endpoint; nothing else changes on the client side.
Per-model routing is configurable via environment variables: a fallback MODEL with optional MODEL_OPUS, MODEL_SONNET, and MODEL_HAIKU overrides, each pointing to whatever backend and model identifier you want. This lets you direct Opus-class prompts differently from the constant lightweight queries Claude Code generates during normal operation.
The proxy also short-circuits trivial Claude Code probe requests locally before they reach any provider. Claude Code generates a nontrivial volume of status-check and lightweight probe calls that don’t require remote inference, and intercepting those matters when the upstream is rate-limited. The proxy additionally handles streaming, tool use, and reasoning/thinking block normalization across backends that don’t all expose extended-thinking equivalents natively.
The Six Backend Trade-offs
The six supported backends split into hosted and local. Hosted options impose provider-specific rate limits and external billing relationships. Local options trade those constraints for hardware-bound inference and no external network dependencies.
| Backend | Type | Key characteristic |
|---|---|---|
| NVIDIA NIM | Hosted, free tier available | Rate-limited per account tier |
| OpenRouter | Hosted, pay-per-token | Aggregated model catalog |
| DeepSeek | Hosted, native Anthropic wire format | Format-compatible endpoint reduces translation overhead |
| LM Studio | Local | GUI-managed model server |
| llama.cpp | Local | Direct GGUF model control |
| Ollama | Local | Simplified local model management |
The quality trade-off that matters most for Claude Code workloads: complex multi-file reasoning tasks that Claude Sonnet handles well will produce different output on any non-Anthropic backend, whether hosted or local. The proxy normalizes the API surface; it cannot normalize model capability. For interactive coding workflows with tight quality requirements, the free-tier path is a cost-reduction experiment, not a drop-in substitution.
What Anthropic’s Policy Whiplash Means for Users
The gap between the April 4 crackdown and the April 21 reversal spans seventeen days and includes a creator account suspension that was publicly reversed within hours. For practitioners building on top of Claude Code, this is a data point on how quickly the acceptable-use surface can move.
The proxy exists outside the specific subscription-vs-API dispute, but not outside Anthropic’s broader interest in controlling how Claude Code is deployed. The April 21 statement endorsed API key access; it didn’t endorse rerouting those calls to competing providers. That distinction matters as the policy continues to develop.
The Proxy Pattern: Tolerance, Fingerprinting, or Client Pinning
Anthropic has three plausible responses to a proxy at this layer.
Tolerance. The proxy routes API-key-billed traffic. If users are paying per token, Anthropic’s systems are not stressed the way subscription-powered harnesses stressed them. The case for intervention is weaker here than it was for OpenClaw.
Fingerprinting at the API layer. Claude Code sends distinctive request patterns: specific tool schemas, system prompt structure, characteristic context window sizing. A request fingerprinting layer could identify Claude Code traffic and reject requests where the model identifier resolves to a non-Anthropic backend. Implementable, but it would also affect legitimate multi-provider setups and produce the same kind of backlash the April 10 Steinberger suspension generated.
Client pinning. If Claude Code required a certificate chain or signed token that only Anthropic issues, proxy interception breaks at the TLS layer. Anthropic has not signaled this, and the April 21 framing (“API keys are the clearest path”) reads as an endorsement of the open API surface rather than a move toward locking it.
The free-claude-code repository1 reports going from 3 HN points in February to 16.9k stars by late April (figures are from the repository itself and have not been independently verified). That trajectory is a legible signal about demand for API cost reduction in Claude Code workflows, and Anthropic will eventually need a policy position on it that doesn’t reverse in seventeen days.
Frequently Asked Questions
Does free-claude-code work with other AI coding tools like Cursor or Windsurf?
The proxy is built specifically for Claude Code’s request format and Anthropic API surface. Cursor, Windsurf, and similar tools use different API contracts — often OpenAI-compatible endpoints — and wouldn’t route through this proxy without modification. The architecture depends on Claude Code’s assumption that it’s communicating with api.anthropic.com.
What caused the adoption spike after the February HN post got zero traction?
The proxy remained obscure from its February 6 posting until Anthropic’s April 4 subscription ban displaced over 135,000 OpenClaw users searching for alternatives. No major tech publication covered free-claude-code directly — TechCrunch and The Next Web covered the OpenClaw policy fight, but the proxy’s 16.9k-star growth came from community discovery among affected users, not press amplification.
Which Claude Code workflows degrade most on a local model backend?
Multi-file refactoring and complex debugging that rely on extended thinking traces degrade first — most local models don’t emit reasoning tokens, so the proxy’s thinking-block normalization either drops the traces or synthesizes a stub. Single-file edits and lightweight queries hold up better because they depend less on extended reasoning. The proxy’s earliest commits, dated February 14, 2026, already included local backend support, suggesting these capability gaps were anticipated from the start.
What happens if NVIDIA discontinues the NIM free tier?
Three of the six supported backends (Ollama, LM Studio, llama.cpp) run entirely locally with no provider dependency, so the proxy itself survives. The cost calculus shifts: users chasing free hosted inference would need local GPU hardware sufficient for code-assistance workloads or would fall back to pay-per-token providers like OpenRouter and DeepSeek, which are typically cheaper per token than Anthropic’s direct API pricing for comparable volume.