375 AI Providers, 153 Free—This Open-Source Gateway Connects Them All. But Where Does Your Code Go?

Someone built a free gateway that connects 375 AI providers (OpenAI, Anthropic, DeepSeek, and the like) to the same endpoint. Of these, 153 are free, with the aggregated free-tier quota estimated at about 1.62 billion tokens per month.

The project is called OmniRoute, MIT licensed, with 63,000 GitHub stars and more than 900 contributors. It runs on your own machine; AI coding tools such as Claude Code, Codex, Cursor, and Copilot send requests to it, and it decides which model provider to forward them to.

The GitHub repository directory page for OmniRoute.

When one provider hits its quota, it automatically switches to the next. Kimi, GLM, DeepSeek, MiniMax, Claude, and GPT are all behind this single endpoint. Requests are compressed before forwarding; official figures say token usage drops by 15% to 95%, with RTK plus Caveman layered together saving an average of 89.2%, while code blocks and JSON structures are preserved as-is.

This works because cheap models are genuinely good enough. In another article, MIKE gave a comparison: in Artificial Analysis's Coding Agent Index 2.0, Claude Code running Opus 5.5 costs $11.42 per task; Codex running DeepSeek V4.1 Flash costs only $0.39. That's 96.6% cheaper. The budget for one Claude task can run about 29 DeepSeek tasks.

That article covered six models worth watching: MiniMax M3.1 Flash, $0.15 per million input tokens, 1M context; DeepSeek V4.1 Flash, top of Code Arena WebDev among open-source models, MIT licensed; Kimi K3.1 Pro, suited to hard reasoning and long documents, open weights (downloadable for self-hosting), released November 4; GLM-5.3-Flash, fast when self-hosted; Ming-3.0, a 0.5B-parameter text-to-speech model, MIT licensed; and MiniCPM-V 4.7, which reads screenshots on a phone and runs at 13+ tokens/s on an iPhone 17 Pro Max. OmniRoute connects all of these into your existing tools. A token is the basic unit of AI billing, roughly equal to most of an English word; 1M context means the model can read text equivalent to several thick books at once.

Using it isn't complicated:

npm install -g omniroute
omniroute

Then, in Claude Code or Cursor, point the Base URL to http://localhost:20128/v1 and set the model to auto. Automatic routing scores candidate providers by health, quota, cost, and latency, then picks a suitable one. When quota runs out, it automatically falls back.

People in the comments asked a few practical questions.

Coding-agent prompts are often entire files plus diffs. Whether a free provider logs them or uses them for training is not something you control. OmniRoute's approach is local-first: keys are stored encrypted with AES-256-GCM, guardrails can filter sensitive information, and providers marked tos: avoid in the directory do not participate in automatic routing by default. But "not by default" does not mean "impossible"; you need to read the specific rules yourself.

There's also the consistency problem. If the model switches mid-task, the new model behaves differently, and the code may look like two people wrote it. OmniRoute has 20 routing strategies; LKGP (last-known-good-provider) tries to stick with the provider that last succeeded, reducing mid-task handoffs. But free tiers are inherently unstable, so you have to assess that risk yourself.

Client projects are another matter. One commenter said plainly that client code cannot flow to 150 random free providers. If the data is sensitive, the safer approach is to download open-source model weights and run them on your own machine—for example, GLM-5.3-Flash or a quantized version of DeepSeek (a compressed version that can run on consumer GPUs); a gaming GPU with 12GB of VRAM can handle it.

The free quota itself also changes. OmniRoute's free directory is re-reviewed every two weeks; when a provider discontinues its free tier, the number goes down. The project itself states clearly: it is not a guarantee, but the result actually calculated from the directory.

Send 90% of everyday work to cheap models and reserve the hardest 10% for Claude. There's nothing wrong with that strategy itself. OmniRoute has connected the toolchain for it. But before plugging company code into it, figure out where the data goes, how long the free quota will last, and what happens when fallback switches the model.

Repository: https://github.com/diegosouzapw/OmniRoute
Original tweet: https://x.com/mikenevermiss/status/2114822790461161617
MIKE's article: https://x.com/i/article/2114621340133052416