Free LLM gateway

Free LLMs, one endpoint, a whole team of agents.

Dozens of free LLM providers behind one OpenAI- and Anthropic-compatible endpoint, with smart routing and agent teams. 100% free and open source.

DownloadCode on GitHubTry the demo
Platforms
Windows, macOS, Linux, Docker
Stack
TypeScript, Express, React, Electron, Docker
License
MIT
Version
v0.1.4
  • 78providers
  • 63free providers
  • 0keys needed to start

How it works

One endpoint, dozens of free models (MocaLLM)

  1. One endpoint

    MocaLLM exposes OpenAI- and Anthropic-compatible endpoints, so Claude Code, Codex CLI, Cursor, Cline or any OpenAI client just point their base URL at it.

  2. Smart routing

    auto, auto:fast, auto:smart or auto:coding rank the models by quality, speed and live reliability, from 63 free providers plus any paid ones you allow.

  3. Failover

    A 429, a timeout or a bad key sends the request to the next model, with exponential cooldowns, key rotation and local counters that keep you under every free limit.

  4. Streaming

    The response streams to your client and every reply carries an X-Moca-Routed-Via header, so you always know which provider answered.

  5. Teams + catalog

    A coordinator can lead a team of free models that plan, delegate and review. The model catalog refreshes from an Ed25519-signed feed, free for everyone.

About the project

Every AI lab now has a free tier. On its own each one is small; stacked together they are a serious amount of inference. MocaLLM does the stacking for you: one base URL, one API key, automatic failover when a provider is rate-limited, and local counters so you stay under every free cap.

It works out of the box with zero keys and gets much better as you add free keys. Paid providers you already have are used only if your policy says so.

One endpoint, every client

Chat completions with streaming, tools and vision, the Responses API for Codex CLI and the Anthropic Messages format for Claude Code — so Cursor, Cline, Aider, OpenCode and friends just work.

Always up to date

Every change is tested and released automatically, installed apps update themselves, and the model catalog refreshes from an Ed25519-signed feed every few hours.

Interactive demo

Simulate a request

Send a request through the router and watch it pick a model and fail over to the next provider when one is rate-limited.

Strategy (the model field)

Balance of quality, speed and reliability.

Paid providers

Tip: press Ctrl + Enter to send.

POST /v1/chat/completions
{ "model": "auto",
  "messages": [{ "role": "user", … }] }
Router orderscore · requests/min
  1. 1Google AI Studiofreegemini-2.5-flash0.840/15
  2. 2Cerebrasfreegpt-oss-120b0.820/30
  3. 3Groqfreellama-3.3-70b0.810/30
  4. 4Mistralfreemistral-small0.790/60
  5. 5NVIDIA NIMfreenemotron-70b0.750/40
  6. 6OVHcloudfreellama-3.1-70b0.730/12
  7. 7LLM7freegpt-4.1-nano0.710/10
  8. 8OpenRouterfreeqwen3-coder:free0.700/20
  9. –OpenAIpaidexcluded by policy0.880/500
  10. –Anthropicpaidexcluded by policy0.870/500
Attempts

Press Send. Some free providers answer 429 or time out: the router moves on to the next one by itself.

In-browser demo with simulated data. The real app does much more.

Features

01

Strictly free by default

The free catalog lists only free tiers. For paid providers you choose the rule: never, fallback only, or prefer.

02

One endpoint, any client

OpenAI and Anthropic compatible, so it works with Claude Code, Codex CLI, Cursor, Cline, Aider and more.

03

Smart routing

Scores models by quality, speed and live reliability, and fails over on 429s, errors and timeouts.

04

Central coordinator

Let Claude Code, Codex or any model lead teams of free models: plan, delegate, review, assemble.

05

Agent teams with skills

8 ready-made teams, 20 agents and 31 skills. Agents search the web, read pages and write files.

06

Always up to date

Self-updating app and a signed model catalog refreshed daily.

07

Private

Keys encrypted on disk, server bound to localhost, prompts not logged unless you turn it on.

08

Bilingual dashboard

English and Romanian, light and dark.

Where it runs

  • Desktop app for Windows, macOS and Linux that lives in your tray.
  • Docker image for a server or NAS.
  • From source with Node.js 20 or newer.

Questions about MocaLLM

Is MocaLLM really free?

Yes. There is no premium tier, no account and no telemetry. The live model catalog is free for everyone.

Can I use it with Claude Code?

Yes. Point ANTHROPIC_BASE_URL at your local MocaLLM and Claude Code runs on free models, with MocaLLM translating formats.

What happens when a free provider hits its limit?

The router fails over to the next best model, puts the limited provider on an exponential cool-down and rotates keys.

Contact

Got a project in mind?
Let's talk.

Tell me briefly what you want to build. I'll reply with a concrete proposal: what goes into the first version, how long it takes and what the steps look like.

Find me on GitHubGenerate a brief