FreeLLMAPI is an open-source repo (almost 50,000 GitHub stars) that stacks the small free plans dozens of AI companies already give developers, 34 providers, 635 free models, behind one unified key. A local router on your own machine picks the best available free model for each request. Claude Code connects to it the same way it connects to Anthropic. • The real number, not the headline: 7.4 billion is every free plan from all 34 providers added up, and nobody holds keys to all 34. With the 4 easiest-to-get keys (Google AI Studio, Groq, NVIDIA, OpenRouter), the app accounts for about 141 million tokens a month, roughly 2% of the headline, still plenty for side projects. • One live request, shown on screen: a single "hi" used about 12,000 tokens in, after Google and Groq both failed first and it fell back to an OpenRouter model. A real test job (build a messy CSV, clean it with Python, run it) took 15 requests total. • The keep rule for which free models are worth connecting: tools support (so it can touch files and run commands, not just chat) and at least 128,000 tokens of context, since Claude Code sends roughly 9,000 tokens of its own instructions before your actual question even arrives. My take: the best free models have the smallest daily limits, Gemini's Flash models get about 20 requests a day each, so that's roughly one job like the demo per model, per day. When the strong models run dry, you're left with the weaker fallbacks. This is a real backup for side projects and learning Claude Code without paying, not a swap-in for client work or anything that has to be right. 📎 Full guide PDF pinned below: the real demo numbers, how the router scores and falls back between models, which specific models got kept on each provider and why, the two-line setup that keeps your paid Claude account untouched, and the repo link itself (it looks like it didn't make it into the YouTube description, a few people are asking for it in the comments): github.com/tashfeenahmed/freellmapi