A single free LLM tier is close to useless: a few million tokens, a few thousand requests a day, one more SDK to babysit. FreeLLMAPI's bet is that eighteen of them, stacked behind one router that never forgets who is rate-limited, add up to roughly 1.7 billion tokens a month of working inference. All of it sits behind a single OpenAI-compatible /v1 endpoint, so any OpenAI client points at your local server and routes across whichever providers you have added keys for.
The provider list reads like a census of every lab currently giving inference away: Google, Groq, Cerebras, NVIDIA, Mistral, OpenRouter, GitHub Models, Cohere, Cloudflare, HuggingFace, Z.ai, Ollama Cloud, and a long tail of anonymous-friendly gateways like Pollinations, LLM7, OVH AI Endpoints, and AI Horde. You can also register a custom OpenAI-compatible endpoint for chat, embeddings, images, or audio, which covers llama.cpp, LM Studio, vLLM, or a remote gateway. Keys are stored encrypted.
The accounting is the actual product
Collecting free tiers is the easy half. The hard half is spending them without tripping a cap, and that is where FreeLLMAPI puts its engineering. It tracks four counters, RPM, RPD, TPM, and TPD, per (platform, model, key), and only routes to a key that is currently under all of them. When a provider returns a 429, a 5xx, or times out, the router skips it, drops the key onto a short cooldown, and retries the next model in your fallback chain, up to 20 attempts against a wall-clock budget. A dead key rotates to its siblings rather than failing the request, and an exhaustion error hands back the full trail of what it tried.
The rate ledger also learns. If a provider names a real limit in an error body or a quota header, say a Groq 413 that reports its TPM ceiling, the router tightens its own numbers to match. Chain order comes from one of six strategies: priority, balanced, smartest, fastest, reliable, or a custom weight mix, scored from live measurements of speed, capability, rate-limit headroom, and recent errors, with a Thompson-sampling bandit choosing under the hood.
One base_url, four kinds of client
The compatibility surface is wider than "the OpenAI SDK works." POST /v1/chat/completions and GET /v1/models cover the OpenAI libraries and clients like LangChain, LlamaIndex, and Continue. On top of that sit three translating shims over the same router. POST /v1/responses speaks the Responses wire format current Codex CLI versions need. POST /v1/completions turns legacy prompt/suffix calls into ghost-text autocomplete for editors. POST /v1/messages speaks Anthropic's format, so Claude Code and the Anthropic SDKs run against your free pool, with /v1/models content-negotiated to return the Anthropic shape when the client sends an anthropic-version header and the OpenAI shape otherwise.
Two features are worth calling out. Fusion is a virtual model: request it and the router fans your prompt to a panel of diverse free models in parallel, then a judge model synthesizes one answer from the drafts, each sub-call still going through normal routing and quotas. Sticky sessions pin a multi-turn conversation to one model for 30 minutes, on the theory that switching models mid-conversation drives a hallucination spike.
Read the disclaimer
The project is candid about what it is. Its own description says personal experimentation only, and the provider table flags the gray areas directly, such as NVIDIA's eval-only terms and AI Horde being anonymous but slow. The model catalog updates itself twice a day from a signed feed, but free installs get each new model on a 30-day delay. The day-one catalog is a $19-a-year Premium tier that the README says is 79 models ahead right now. That funding pitch sits in plain sight next to an MIT license and a TypeScript codebase you can run yourself. Whether stacking a dozen free tiers this way stays inside every provider's terms is a question the repo raises and leaves to you, which is the right place to leave it.
GitHub: https://github.com/tashfeenahmed/freellmapi
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)