Short version: FreeLLMAPI stacks the free tiers of 87 AI providers behind one OpenAI-compatible endpoint, and the README promises 10.1 billion free tokens a month. I read the code instead of taking the number at face value. The number is real arithmetic, and it is also not what you will actually use. The interesting parts are how the number is made, what a day of real use looks like, and the fact that the project tells you, in writing, where you should not use it.
What it actually is
FreeLLMAPI is a self-hosted router, written in TypeScript. You bring your own free-tier API keys, from Google, Groq, Cerebras, Mistral, OpenRouter, and eighty-odd others, and it gives you one OpenAI-compatible /v1 endpoint to point your tools at. The router picks a model, retries the next provider when one rate-limits you, and tracks per-key usage so you stay under each free-tier cap. It also speaks the Anthropic messages format and the Gemini API, so Claude Code and Gemini CLI work against the same pool.
Your keys are stored encrypted in a local SQLite database, AES-256-GCM, decrypted in memory per request. Your applications only ever see a single unified key. It runs as a desktop app, a Docker container, or plain npm, on port 3001, at about 40 MB of memory idle. The install is the classic one-liner,
curl -fsSL https://freellmapi.co/install.sh | bashand to the README’s credit it links the script and tells you to read it first.
How the 10.1 billion is made
Here is the part I liked most, because it is checkable. The model catalog stores each free tier as a human-written label, something like ~120M or ~50-100M. The router parses the upper bound of the label. A daily cap gets multiplied by 30 to make it monthly. Add up all 1,313 endpoints and you get 10.1 billion.
So the number assumes you burn every endpoint’s full allowance, every month, including the ones you will never touch. It is listed capacity, not usable capacity. The breakdown is 1,137 chat endpoints, 29 image, 41 audio, 68 embeddings, 19 transcription, 6 video, and 13 decision models. That adds up to 1,313, and I checked.
The code comments are candid about how fragile that parsing is. An earlier version of the parser read “40 RPM” as 40 million tokens, because the M in RPM looked like a magnitude. It is fixed now, and the fix is documented in the code, with the bug described in plain language. I like a project whose source code documents its own measurement errors. It makes the headline number easier to trust, because you can see exactly how it is built and where it could be wrong.
What a day of real use looks like
The repo’s own Limitations section is the honest version of the pitch, and it is worth quoting in spirit: the catalog does carry frontier-class models, GPT-5.x, Grok 4.x, Kimi K3, DeepSeek V4 Pro, Gemini 3.x, but those are exactly the rows with the smallest daily allowances and the highest chance of being pulled or paywalled at short notice.
The top models hit their caps first. When they do, the router falls down your priority chain to smaller, weaker models. The effective intelligence of the endpoint drops as the day goes on, then resets at UTC midnight. Latency varies wildly depending on which provider answers. There is no SLA, because there cannot be one on somebody else’s free tier.
None of that is a flaw in the router. It is the shape of the thing being routed. But it is the difference between 10.1 billion tokens and 10.1 billion tokens you would want.
The catalog is the product
This is the part most coverage misses. The router is MIT licensed and free forever. What costs money is freshness.
Twice a day, your router pulls a signed model catalog from the project’s server. The download is verified against a pinned Ed25519 public key baked into the code, so a tampered feed gets discarded. Free installs get the monthly snapshot, which means a new model reaches them 30 days after it joins the live feed. The README says free builds currently sit about 303 models behind. Premium is $19 a year or $49 once, and gets the same feed the day a change ships.
That is open core done the honest way. The paid part is how fast you hear about a change, not a feature gate. Nothing expires, nothing is crippled, updates just arrive later.
Two details from reading the code: the repo ships about 56 provider adapter files, so the 87 number is the catalog’s count, including providers that only exist in the feed. And I searched the server code for telemetry and found none. The only calls back to the project are the catalog sync and the update check.
The terms of service table
The project re-reviewed each provider’s terms of service in May 2026 and published the verdicts in its docs. Cohere is marked avoid, because its terms forbid personal, family, or household use. Google gets a caution flag, because a March 2026 clause narrows its free tier to professional or business purposes. NVIDIA’s free tier is evaluation only, not production. Z.ai gets a caution for an anti-traffic-redirect clause that could plausibly be read against a proxy. The rest are likely fine.
The rules of thumb are short enough to remember: one account per provider, no reselling, no sharing your endpoint with other people, and don’t hammer a free tier as a production backend. The disclaimer is equally short: this is for personal experimentation and learning, not production, and the providers’ terms still apply when the traffic is proxied through the project.
A project that tells you where not to use it is a project you can take seriously. Most aggregation pitches hide exactly this part.
What I did not do
I did not run it. I read the code, the docs, and the catalog logic, but I have not installed it or pointed an agent at it, and this post is not a review of running it.
The test that would settle the headline is simple: install it, add three or four free keys, run a real coding session through it for a day, and measure tokens served, what that would have cost elsewhere, and where it falls over. That is the next post, if I run it.
The 10.1 billion is a sum, not a promise. What is real is the router, the catalog, and a project that documents its own limits better than most paid products document their features. Free tiers are a commons with a fence around it. This is a map of the fence.
