FreeLLMAPI

FreeLLMAPI is an MIT self-hosted gateway that puts dozens of free LLM tiers behind one OpenAI-compatible endpoint, with routing and failover.

Read this first: it is for personal experiments

• The README's own disclaimer says the project is for personal experimentation and learning, not production. Free tiers are not a stable or supported inference service, so if you build something real on it, swap in a paid API before you ship.
• Your terms with each upstream provider still apply when traffic goes through the proxy, and you are responsible for following them. The docs include a Terms of Service review from May 2026 for a self-hosted, single-user setup. It marks Groq, Cerebras, Mistral and OpenRouter as likely fine, Google Gemini, NVIDIA NIM and Z.ai as caution, and Cohere as avoid. The author wrote it, and it is not legal advice.
• It is local-first, with no multi-tenant auth. The docs say to run it for yourself and not to expose it to the internet.
• The one-line installer pipes a script into bash. The README links the script so you can read it first.

What it is: many free tiers behind one key

The repo is by tashfeenahmed, MIT licensed, with a site at freellmapi.co. You add your own keys for each provider, and your apps talk to one local endpoint with one unified key. A router picks a model for each request, moves to the next provider when one is rate-limited, and counts usage per key to stay under each free-tier cap. Provider keys are stored encrypted with AES-256-GCM.

• The README says the catalog covers 34 free providers, 474 model families and 635 free endpoints, adding up to about 7.4 billion tokens a month of listed free capacity. These are the project's figures, and free tiers change without notice.
• The API covers chat completions, responses, completions for editor autocomplete, embeddings, image, video and speech generation, transcription and model lists. It also speaks Anthropic's Messages format, and has Gemini and Ollama-style surfaces. You can add your own OpenAI-compatible endpoint, such as llama.cpp, LM Studio, vLLM or a local Ollama.
• Routing has six strategies, with named fallback chains such as a coding chain or a vision chain, sticky sessions of 30 minutes, tool calling that routes only to models that support it, and an opt-in "fusion" model that asks several free models and has a judge model combine the answers.
• The dashboard manages keys, reorders the chain, has a playground and shows latency analytics. There is also an MCP server, an opt-in response cache and opt-in prompt compression.

Using it with coding agents

Point an OpenAI client at http://localhost:3001/v1 with the unified key. The README lists Claude Code, Codex CLI, Gemini CLI, Cursor, Cline, Roo Code, OpenCode and Aider, plus Continue, Goose, Qwen Code, Kilo Code, Crush, Zed, JetBrains AI, DeepSeek Harness and Hermes Agent, among others. Most are set up with one command that reads your live catalog and backs up the existing config, for example npx freellmapi setup-claude --url http://localhost:3001 --api-key <unified-key>.

Install

• curl -fsSL https://freellmapi.co/install.sh | bash needs Docker. It sets up ~/freellmapi, makes an encryption key, pulls the image and starts it. Open localhost:3001, add your provider keys on the Keys page and copy the unified key.
• Docker Compose, a source run with Node 20 or newer, and a desktop menu-bar app with macOS (Apple Silicon and Intel, macOS 12 or later) and Windows installers on the Releases page. An Android install through Termux is marked experimental.

Free and paid

• The README says the router stays MIT and free. Free installs get the signed model catalog from a monthly snapshot, so a new model arrives 30 days after it reaches the live feed. At the time of the README, it said the free build sat about 303 models behind.
• The paid option is the live catalog: $19 a year or $49 once, with the same-day updates on every router you run. The README says this funds the daily model testing, and that the catalog server never sees your prompts, completions or provider keys.
• Check each provider's own limits and billing. The v0.16.20 release notes say the FreeModel provider needs your own key and Telegram verification, its 1 million tokens a month are shared across models, and a funded wallet can be charged after the allowance, so keep it unfunded for free-only use.

Limits the author lists

• Frontier-class models appear on some free tiers, but with the smallest daily allowances and the highest chance of being pulled or put behind a paywall. The author says to budget for capacity and uptime, not for a capability limit.
• Quality drops through the day. Your best models hit their daily caps first, so the router falls to smaller ones until the reset at UTC midnight. Latency varies by provider, and there is no SLA.
• Not supported yet: moderation, n greater than 1, and per-user billing.

How fresh it is

• The repository was created on April 21, 2026. The latest release, v0.16.20, came on November 6, and the last push was on November 7.
• On November 7 it had about 48,500 stars, about 6,700 forks and 47 open issues and pull requests together.
• Seailife did not install it, add provider keys or send a request through it. The counts, provider terms and setup steps above are from the README and the docs.