OmniRoute
by Diego Souza and contributors
“Never stop coding.” A free MIT-licensed AI gateway: one endpoint fronting hundreds of providers and thousands of models, with quota-aware auto-fallback so a rate limit on one provider silently rolls over to the next.
See https://omniroute.online · source at https://github.com/diegosouzapw/OmniRoute
Licence — confirmed
MIT, Copyright (c) 2026 diegosouzapw. Verified in both LICENSE and the "license": "MIT" field in package.json, despite GitHub’s API reporting None. The repo also ships a 303-line THIRD_PARTY_NOTICES.md, which is better dependency hygiene than most projects of its size manage.
What it is
A gateway in the LiteLLM/OpenRouter mould: your coding agent points at one OpenAI-compatible endpoint, and OmniRoute handles which upstream provider actually serves the request. Its specific angle is free tiers. It catalogs free-tier allowances across providers and routes to keep you inside them, so the pitch is uninterrupted coding at zero marginal cost.
Works with Claude Code, Codex, Cursor, OpenCode, Cline, Copilot and Antigravity. Ships Docker and docker-compose, a Fly.io config, a Nix flake, an Electron desktop build and a PWA.
Numbers, and how to read them
The README leads with large figures — 351 providers, 1,200+ models, 90+ free tiers, ~1.51B free tokens/month, 19 routing strategies. What makes them more trustworthy than the usual README maths is that the project shows its method:
- The token headline is computed from 20 pools with a published positive monthly budget, deduplicated by shared pool, out of 445 cataloged free-tier entries across 39 recurring pool keys. Methodology is written up in
docs/reference/FREE_TIERS.mdand surfaced live at/dashboard/free-tiers. - Figures are re-audited every two weeks and explicitly “move both ways” — the README states that when a provider ends a free tier the number drops, and that they publish what the catalog computes rather than a rounded-up best case.
- A terms-risk catalog marks 13 providers “avoid” so you can decide, rather than silently routing you through them.
That is unusually honest for a project marketing free capacity, and it is the main reason to take the claims seriously.
Features
- Quota-aware auto-fallback — rolls over when a provider’s limit is hit, across 19 routing strategies
- Token compression — “RTK + Caveman” stacked compression claimed at 15–95% savings (~89% average)
- MCP and A2A support
- Desktop (Electron) and PWA front ends, plus a dashboard
- 47 bundled skills
- Deployment via Docker, docker-compose, Fly.io or Nix
Engineering signals
The repo is more rigorously set up than its README tone suggests: Vitest with separate configs for e2e-live and MCP, Playwright, Stryker mutation testing, SonarQube, Codecov, promptfoo for prompt evaluation, and four ESLint configs including dedicated complexity ratchets. Mutation testing and complexity ratchets in particular are not things you find in a vibe-coded project.
Caveats
- The README is extremely promotional — dozens of badges, multiple WhatsApp and Telegram groups, repeated star solicitations, affiliate-shaped links. Judge the code, which is stronger than the packaging.
- Free-tier routing is a moving target by the project’s own admission; capacity you plan around can disappear at a provider’s discretion.
- Provider terms are your risk. The “avoid” list helps, but funnelling work through many free tiers can conflict with individual providers’ terms of service. The catalog surfaces this rather than resolving it for you.
- Compression changes your prompts. 15–95% token reduction means something is rewriting what reaches the model; verify quality on your own workload before trusting the savings.
- “Built by 450+ contributors” is a headline figure worth treating like the others — impressive, but contributor counts include drive-by commits.
Pricing
Free and open source under MIT, self-hosted. The point of the project is to minimise provider spend rather than to charge for itself.
Related
- Model Context Protocol — supported alongside A2A
- Claude Code · Codex · Cursor · OpenCode — clients it fronts
- Claude Code Router — narrower tool solving the same “point Claude Code elsewhere” problem
- herdr · WeSight · Orca — orchestrate agents; OmniRoute instead sits underneath them at the model layer