FreeLLMAPI routes 34 free LLM tiers behind one OpenAI-compatible endpoint

An open-source project called FreeLLMAPI offers a way to combine the free tiers of 34 LLM providers behind one OpenAI-compatible API running on your own machine. According to a post on dev.to, the router accepts any OpenAI client pointed at a local /v1 endpoint and distributes requests across whichever providers you have supplied keys for, with Google AI Studio, Groq, Cerebras, Mistral, Cohere, NVIDIA, Cloudflare and OpenRouter among the supported options.

The problem it targets

The dev.to post frames the motivation plainly: nearly every AI lab now runs a free tier, and each individual allowance is modest — a few million tokens per month or a few thousand requests per day. Taken together, however, they add up to enough capacity for genuine prototyping. The obstacle is integration: separate SDKs, separate rate limits and separate failure points for every provider you add. FreeLLMAPI's answer is to collapse that complexity into a single local component.

What the router does

On each request, the router consults your fallback chain and selects the highest-priority model that has a healthy key and sits under all of its rate limits. If a provider answers with a 429 or a 5xx, the key is placed on cooldown and the next model in the chain is tried immediately. The system tracks requests-per-minute, requests-per-day and the corresponding token counters for each provider, model and key combination, aiming to stay under caps rather than discover them by rejection. An X-Routed-Via response header records which provider actually served the call.

Several conveniences sit on top of the core routing. Sticky sessions hold a conversation on one model for 30 minutes to keep replies coherent, the same model offered by multiple providers appears as a unified entry, and named routing profiles can be selected per request. The model field accepts "auto" to let the router choose, "auto:fast" or "auto:smart" to bias toward speed or capability, an explicit model id, or "fusion", which sends the prompt to several free models in parallel and has a judge model merge the results into one answer.

Local-first key design

There is no hosted signup. You collect free-tier keys yourself and paste them into a dashboard, where they are encrypted with AES-256-GCM into a local SQLite database and decrypted in memory only for the life of a request. The router then issues a single bearer token prefixed with freellmapi-, which is the only credential your application ever handles. The project does not resell tokens and, because it runs locally, does not pass your prompts through a remote server of its own.

Setup options include a Docker install script, npm, and desktop installers for Windows and macOS that run from the system tray. The dashboard listens on localhost:3001, keys appear with health-check status indicators, and the fallback order is arranged by dragging models into place on a dedicated page.

What is free, and what is not

The software is MIT-licensed and every feature works at no cost, the post says. The single paid item is a live model catalog feed, priced at $19 per year or $49 for a lifetime licence. Free installations receive the same signed catalog as a monthly snapshot, so newly launched models arrive roughly 30 days later; the project states that free builds currently trail the live feed by about 303 models. Nothing is disabled on free installs — updates simply arrive later.

Why it matters

Free inference capacity already exists across the industry, but the overhead of juggling many providers has kept it fragmented. Wrapping pooled free tiers in a standard OpenAI-compatible interface means existing tooling — including CLI coding agents, which the project says can be configured with one command — can draw on that capacity without code changes.

The limits are stated openly. This targets personal experimentation and learning, not production: there are no frontier models, no SLA and variable latency. The dev.to post also notes that effective quality declines late in the day as popular free models exhaust their daily caps, which reset at UTC midnight, and that your use of each provider remains governed by the terms you accepted when signing up. Since the details here come from the project's own announcement, claims such as the encryption scheme and model counts would benefit from independent verification. For anything beyond prototyping, the intended path is to swap in a paid API.