FreeLLMAPI: 7.4 Billion Free AI Tokens a Month From 34 Providers Behind One Endpoint

FreeLLMAPI is a self-hosted router that stacks the free tiers of 34 AI providers behind a single OpenAI-compatible /v1 endpoint. Every serious AI lab offers a free tier, a few million tokens a month or a few thousand requests a day, and on its own each one is a toy. Stacked together they add up to roughly 7.4 billion tokens a month across 474 model families and 635 free endpoints, from small and fast models to reasonably capable ones. Point any OpenAI client library at your local server and it routes across whichever providers you added keys for. The project is written in TypeScript, released under the MIT license, and it has 50,376 stars.

The router picks the best available model for each request and falls over to the next provider when one is rate-limited. It scores every model on live speed, capability and reliability, then retries the next model in your chain on a 429 or 5xx error, with cooldowns and key rotation. Per-key rate tracking keeps RPM, RPD, TPM and TPD counters for each platform, model and key, so routing stays under every free-tier cap. If a provider feels slow from your side, a Speed Test tells you whether your own connection is the bottleneck before you blame the router. The same model on several providers collapses into one entry with strict in-group failover, and named fallback-chain profiles let you keep a coding chain and a vision chain and switch between them per request.

FreeLLMAPI covers every OpenAI-style surface: chat completions, responses, completions for editor ghost-text autocomplete, image generation, video generation, audio speech and transcriptions, embeddings and models, streaming and non-streaming. There is also an Anthropic Messages API surface, so Claude Code and the official Anthropic SDKs run against your free pool, plus a native Gemini surface for Gemini CLI and opt-in Ollama emulation for Zed, JetBrains and other local-model clients. A Fusion mode fans one prompt out to a panel of diverse free models in parallel and has a judge model synthesize one answer from the drafts. Tool calling and structured outputs round-trip across providers, so coding agents get real tool calls rather than plain text.

Most agents configure themselves with one command that reads your live catalog, backs up the existing config and merges into it. For Claude Code it is npx freellmapi setup-claude with your server URL and unified key, and there are matching commands for Codex CLI, Gemini CLI, Cursor, Cline, Roo Code, OpenCode, Aider and more, plus Continue, Goose, Qwen Code, Kilo Code, Crush, Zed, JetBrains AI and any OpenAI-compatible, Anthropic, Gemini or Ollama client. Conversations stay on one model for 30 minutes, and an optional compact handoff note keeps the thread coherent when a mid-chat switch happens. The opt-in prompt compression pipeline can deduplicate prompts, filter tool output and trim stale context before routing, and running your prompt text through an AI Text Cleaner first is another easy way to cut wasted tokens before a request ever leaves your machine.

The fastest start is the one-liner that sets up ~/freellmapi, generates an encryption key, pulls the Docker image and starts the container, then you open http://localhost:3001, add provider keys on the Keys page, reorder the fallback chain and grab your unified key from the header. That unified key is what you point your OpenAI SDK at, with base URL http://localhost:3001/v1 and model "auto" to let the router pick. There is also a native menu-bar desktop app for macOS and Windows, with the whole router and dashboard running locally from the tray, plus Android via Termux and a Raspberry Pi friendly footprint of about 40 MB RAM at idle. Provider keys are AES-256-GCM encrypted in SQLite and decrypted in memory per request, and your apps only ever see the single unified bearer token. The dashboard ships in 60 languages and shows p50/p95/TTFT analytics over 24 hour to 90 day windows.

FreeLLMAPI describes itself as built for personal experimentation, and the limits are real: no frontier models, variable latency, no SLA, and the effective intelligence of the endpoint dips late in the day as top models hit their daily caps, then resets at UTC midnight. Each provider's free tier has its own rules and limits, so the 7.4 billion figure is the sum of the listed free-tier capacity, not a guaranteed allowance for any one user. The router keeps its model catalog fresh on its own, pulling a signed catalog twice a day, but free installs track the monthly snapshot, so a model reaches them 30 days after it joins the live feed. The live catalog is a $19 per year premium add-on (or $49 once, lifetime), and it is the only paid part: the router itself stays MIT licensed and fully free, forever.