FreeLLMAPI
tashfeenahmed/freellmapiReviewed by AI
What it does
FreeLLMAPI runs a small server on your computer that collects the free plans you already have with providers such as Google, Groq, Cerebras and OpenRouter. You paste each key once into its dashboard; your programs then use one local address with one key, and it picks a model that still has quota and moves on when a provider refuses. It works with OpenAI, Anthropic, Gemini and Ollama style clients, including Claude Code and Cursor. Desktop installers, phone apps and Docker. The author labels it for personal experimentation only.
How easy it is to install
B · installer plus one simple step · Windows
- Windows B
- macOS B
- Linux B
Windows is the easiest: the FreeLLMAPI-Setup .exe from the releases page, which the README calls the easiest path, and the desktop app needs no account or password. macOS has a signed-looking .dmg for Apple Silicon and Intel, and Linux has AppImage, .deb and .rpm. It is a B because nothing routes until you have signed up with providers yourself and pasted their keys into the Keys page. The server route (one-line script, Docker Compose) needs Docker and is harder.
Numbers on GitHub
- Stars
- 45,200on Oct 31, 2026
- Downloads
- 396,12334,083 this week
- Growth
- +3,018stars in 5 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux
- You need to install
- Nothing extra
- To build from source
- Node.js, Docker
- Disk space
- 104 MB–311 MB · from the installer
- Accounts or keys
- Free-tier keys from the providers you choose (Google, Groq, Cerebras, Mistral, OpenRouter, Cloudflare, Cohere, NVIDIA, Hugging Face and 25 more), FreeLLMAPI premium key (fla_), for the live catalogue
- Cost
- Freemium · The program is MIT and free. A paid plan ($19/year or $49 once, by Stripe) only unlocks the same-day model catalogue; the free install gets a monthly snapshot and the README says it "currently sits about 303 models behind". The provider free tiers themselves cost nothing but are capped.
- License
- MIT (Permissive)
- Can replace
- OpenRouter, Portkey, ChatGPT Plus
- Last update
- Oct 31, 2026
Risk signals (7)
- Attentionauthor says personal experimentation only
The repository description ends with "Personal experimentation only." and the Limitations section says: "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", with a pointer to "the honest list ... before building anything real on this".
AI review
- Attentiondepends on other companies free tier rules
The whole purpose is to stack the free plans of 34 companies ("7.4 billion tokens per month") behind one address. Each of those free tiers has its own terms, quotas and rules about accounts and automated use, and a provider can tighten a quota, block a key or close an account at any time; the README itself says "the free-tier landscape shifts weekly". The paid catalogue exists because of that.
AI review
- Attentioninstall one liner pipes a script to bash
Quick start: `curl -fsSL https://freellmapi.co/install.sh | bash`, which downloads a script from the project's website and runs it immediately. The README does offer the script for reading first, and the desktop installers avoid it.
AI review
- Attentionkeeps all your provider keys in one database
Provider keys are "AES-256-GCM encrypted in SQLite and decrypted in-memory per request". Outside production the encryption key is generated automatically and written to a file named .encryption-key next to that database with 0600 permissions, so the lock and the box sit in the same folder. The desktop app also signs the dashboard in "with a hidden local account", with no password.
AI review
- Infocontacts the vendor twice a day
The router "pulls a signed catalog from freellmapi.co twice a day", verified against a pinned Ed25519 key. The README says "The catalog server never sees your prompts, completions, or provider keys", and the premium plan is what makes that feed same-day.
AI review
- Infooptional auto update needs the docker socket
docker-compose.yml has an opt-in `autoupdate` profile running Watchtower with `/var/run/docker.sock` mounted, and its own comment says this "is root-equivalent on the host — leave the profile off if that is not acceptable".
AI review
- Infonetwork exposure is the users responsibility
The compose file binds the port to 127.0.0.1 and states FreeLLMAPI "is single-user and must not be exposed to the internet"; HOST_BIND=0.0.0.0 opens it to the local network, "guarded only by the unified API key".
AI review
Evidence is shown in its original language.
What Radar could not check
- Whether the advertised capacity (7.4 billion tokens a month, 635 endpoints) is reachable in practice, and how much of it your own accounts actually get.
- How well the answers hold up when requests bounce between many different free models.
- Whether checking key health spends money: the configuration notes that ModelScope validates a key with a paid one-token request that consumes quota on each probe.
- Whether the desktop installers are signed, and what Windows or macOS warn about when you open them.
- What the free-tier providers log about your prompts on their side.