FreeLLMAPI — a self-hosted router that stacks the free tiers of 34 LLM providers behind a single OpenAI-compatible `/v1` endpoint, with automatic failover and per-key quota tracking.
Alternative to: OpenRouter, which takes no markup on inference but charges 5.5% with a $0.80 minimum on credit purchases (5% via crypto), and whose own free models are capped at 20 requests/minute and 50/day until you have bought $10 of credit. Also the raw API bills it routes around: OpenAI gpt-5-nano at $0.05/$0.40 per MTok, Anthropic Claude Haiku 4.5 at $1/$5 per MTok.
Cost to run: the router is free. Its own live catalog is $19/yr or $49 once. Free installs get a monthly snapshot that trails the live feed by 30 days — currently "about 303 models behind," by the project's own count. Plus your own hardware.
Licence: MIT. Free for commercial use.
Runs on: Docker, macOS, Windows, Android. There is a CLI and a desktop app.
Requires: your own API keys, one per provider. Docker for the standard install.
Traction: 22,940 stars · 3,163 forks · 647 commits · 66 committers in the last year.
Liveness: very live — 182 commits in 30 days, v0.9.0 tagged 2026-08-26.
Maturity: v0.9.0, pre-1.0, and the README's Disclaimer opens: "This project is for personal experimentation and learning, not production."
Links: repo · freellmapi.co · Hacker News, 4 points
Figures as of 2026-08-31.
What it does
Every major AI lab publishes a free tier. Individually each one is a toy — a few million tokens a month, a few thousand requests a day. FreeLLMAPI's argument is that stacked together they add up to real capacity, and that the painful part is not the capacity but the plumbing: 34 different SDKs, 34 different rate limits, 34 places a request can fail.
So it collapses them into one endpoint. You add your own keys, locally, encrypted. A router picks an available model per request, fails over when one is rate-limited, and tracks per-key usage so you stay under each cap. Point any OpenAI-compatible client at your local server and it routes underneath.
It also accepts arbitrary custom OpenAI-compatible endpoints, which is how people wire it into coding agents.
The project's own headline number is roughly 7.4 billion tokens per month across 474 model families and 635 free endpoints. Treat that as the project's count of listed free-tier capacity, not an independently verified figure — the README itself hedges it as "roughly" and "listed," and publishes no methodology.
What it replaces, and what that costs
OpenRouter is the honest incumbent, and it is cheaper than people assume. Its FAQ is explicit: "We pass through the pricing of the underlying providers; there is no markup on inference pricing (however we do charge a fee when purchasing credits)." That fee is 5.5% with a $0.80 minimum on card top-ups, 5% on crypto. There is no subscription.
OpenRouter also has free models — and its limits are the thing FreeLLMAPI exists to escape: 20 requests per minute, and 50 requests per day until you have bought at least $10 of credit lifetime, after which it rises to 1,000/day.
Under both sits the actual price of tokens. OpenAI's cheapest current model runs $0.05 in / $0.40 out per million tokens. Anthropic's Haiku 4.5 is $1 / $5. Google's Gemini Flash tiers have a genuine free column on the pricing table.
So what does FreeLLMAPI save? For a hobbyist prototyping against small models, the honest answer is a few dollars a month, and a lot of plumbing. The plumbing is the real product. Anyone whose workload would cost hundreds of dollars on paid APIs is, by the project's own Disclaimer, not the intended user.
And it has a price of its own. The router is MIT and, in the project's words, "stays MIT-licensed and fully free, forever." What costs money is the model catalog it depends on: $19/year or $49 once. Free installs pull the same signed catalog from a monthly snapshot, so a model reaches them 30 days after it joins the live feed. The README quantifies today's gap at about 303 models behind.
That is worth stating precisely, in both directions. Nothing is crippled and nothing expires — routing, failover and quota tracking are free forever, and the project says so plainly. But the entire value proposition rests on tracking a landscape the README itself describes as shifting weekly, and the free build tracks it a month late. Open core, with the delay applied to the one thing that decays.
Is it still alive?
Very. 182 commits in the last 30 days and 566 in 90, across 66 committers in the past year. v0.9.0 landed 2026-08-26, v0.8.9 and v0.8.8 the day before, v0.8.7 on 2026-08-24. The last commit was 2026-08-28.
The issue tracker is busy and heavily Chinese-language. Open issues from the last week include a Groq 400 on `reasoning_content` when replaying history after a reasoning-model turn, models being force-classified as chat so multimodal categorisation fails, a request for a QwenPaw setup guide, and — the interesting one — #1065, reporting that the monthly token budget bar counts shared pools once per model and displays 2.37B where the actual figure is 294M. A user checking the project's own headline arithmetic and finding it eight times high in the UI is exactly the kind of open issue worth knowing about before you trust a capacity number.
The star count moved from 22,912 to 22,940 during the hours this record was compiled.
GitHub's signed-out view shows an "Issue creation is restricted" banner here too. It is false — outsiders are filing issues daily.
What people say
Almost nothing, in English, and that is a genuine finding for a 22,940-star repository. Its best Hacker News showing is 4 points and zero comments on 2026-08-15; an earlier submission in May got 2 points and no comments. There is no Trendshift placement. Reddit is blocked from this column's research sandbox, so no Reddit vote counts are claimed.
The community that exists is on the issue tracker and it is largely Chinese-speaking. The most on-point thread for a prospective user is issue #1062, titled 有升级高级版的么？好用么？值不值得？ — "Is there a premium upgrade? Is it any good? Is it worth it?" The question people are actually asking about this project is whether to pay for it.
The most credible voice on the project's risks is the project itself, which is unusual enough to note. Its architecture documentation carries a provider-by-provider terms-of-service review, dated May 2026, that rates Cohere "❌ Avoid" because its terms forbid "personal, family or household purposes," flags Google as "⚠️ Caution" over the March 2026 narrowing to "professional or business purposes, not for consumer use," and closes with a rule of thumb: "one account per provider, no reselling, no sharing your endpoint with other humans, don't hammer a free tier as a paid production backend."
That is more candour about legal exposure than most projects in this category manage.
The catch
Read the Disclaimer, because it is the most honest paragraph in the repository. In full: "This project is for personal experimentation and learning, not production. Free tiers exist so developers can prototype against them; they aren't a stable, supported inference substrate and shouldn't be treated as one. If you build something real on top of FreeLLMAPI, swap in a paid API before you ship. Your relationship with each upstream provider is governed by the terms you accepted when you created your account — those terms still apply when the traffic is proxied through this project, and you're responsible for complying with them."
Now the clause nobody quotes. Groq's Acceptable Use Policy, effective 2025-10-15, prohibits customers from using the services "beyond published parameters, rate limits, or use limitations, including by registering multiple accounts or orchestrating usage between multiple organizations." That is the exact behaviour named and prohibited.
Read carefully, this does not condemn the tool. Running your own single key through your own single-user proxy is not prohibited by Groq, or by OpenAI, or by Google, or by Mistral — this column checked all five providers' actual terms and found no clause banning it. What every one of them prohibits is transferring or selling keys, granting third parties access, and deliberately exceeding published limits. FreeLLMAPI's own "one account per provider" rule is precisely what keeps a careful user on the right side of that line. The tool makes it very easy to cross anyway.
Two things the project's own ToS review gets wrong, and both matter:
It rates Cerebras "✅ Likely OK." Cerebras no longer has a free tier at all. Its own documentation, asked whether there is a permanently free tier, answers: "No. The Free Trial is time- and credit-bounded: $5 in credits that expire 30 days after they're granted." And the credits require a verified payment method before API access activates.
It rates Groq "✅ Likely OK" citing only the Services Agreement, without engaging the Acceptable Use Policy clause quoted above.
Google's free tier is paid for with your prompts. Its Gemini API terms are unambiguous: on unpaid services "Google uses the content you submit to the Services and any generated responses to provide, improve, and develop Google products and services and machine learning technologies," and "human reviewers may read, annotate, and process your API input and output." The instruction that follows is Google's own: "Do not submit sensitive, confidential, or personal information to the Unpaid Services." Groq, by contrast, is contractually barred from training on your data.
The quality ceiling is real and the project states it. From the Limitations section: "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."
The headline numbers do not reconcile. The homepage says 605 models; the README says 474 model families and 635 endpoints. Both claim 7.4B tokens/month. Nothing in the repository reconciles the two, and open issue #1065 suggests the in-app budget display is separately inflated.
Credits
Maintained by Tashfeen Ahmed, who wrote 443 of the last year's commits. Contributions from suantea (56 across two accounts), OhOkThisIsFine (18), tashdroid (12), Jasnoor Gill (9) and 61 other committers. Premium is sold by Neu Software LLC via Stripe.
Issue reporters referenced: amansingh909 (#1065), Wan7451 (#1062), silvertakana, yangq626, QRBG3376, growster.
Terms and pricing verified against console.groq.com/docs/legal/ai-policy and its services agreement, ai.google.dev/gemini-api/terms, developers.google.com/terms, openai.com/policies/services-agreement, inference-docs.cerebras.ai/support/rate-limits, legal.mistral.ai, openrouter.ai/docs/faq and its limits page, developers.openai.com/api/docs/pricing, and platform.claude.com/docs/en/about-claude/pricing. Project claims quoted verbatim from the repository's README and docs/architecture.md. Repository figures from a blobless clone and ungh.cc, both read 2026-08-31.