// the find
diegosouzapw/OmniRoute
Never stop coding. Free MIT AI gateway: one endpoint, 290+ providers (90+ free), 500+ models — Kimi, Claude, GPT, OpenAI, Gemini, GLM, DeepSeek, MiniMax. Works with Claude Code, Codex, Cursor, OpenCode, Cline & Copilot. Quota-aware auto-fallback, RTK+Caveman compression saves 15-95% tokens, MCP/A2A, Desktop/PWA. Built by 500+ contributors
OmniRoute is a local proxy that sits in front of your AI coding tools (Claude Code, Cursor, Cline, etc.) and routes requests across 290 providers — 90+ with free tiers — with automatic fallback when a quota or provider dies. It's for developers who want to stop paying full price for AI coding assistants by stacking free tiers and cheap providers behind a single OpenAI-compatible endpoint. Self-hosted, MIT licensed, runs as a local server or Electron app.
The fallback routing is the actual useful part: 19 strategies including quota-aware headroom routing and circuit breakers per provider, per key, and per model independently — so one 429 doesn't kill your session. The RTK+Caveman token compression is a concrete differentiator; compressing tool-call-heavy sessions before they hit the wire is something most routers skip entirely. The free-tier math is audited and CI-gated — a `check:docs-counts` gate fails the build if the headline 1.53B figure drifts from the catalog, which is unusually disciplined for a project of this kind. The opencode-plugin and one-command setup scripts (`setup-claude`, `setup-cursor`, etc.) reduce the integration friction that kills adoption for tools like this.
The scope is genuinely hard to trust at face value: 290 providers, 500+ models, 19 routing strategies, MCP with 104 tools, A2A, memory, guardrails, TLS fingerprint stealth, MITM TPROXY, Electron desktop, PWA, Termux — all in one repo with 500+ contributors. That surface area almost certainly means shallow implementations across most of it, and the README reads like a feature checklist rather than a product. The free-tier reliance is fragile: providers end free tiers constantly, and even with the bi-weekly audit you're building workflows on quicksand. The token compression claim of '~89% avg' is a red flag — that number almost certainly comes from pathological benchmark inputs, not typical coding sessions; real savings for most users will be 15-30%. No mention of latency overhead from local routing, compression, and circuit-breaker bookkeeping, which matters a lot when you're waiting for a code completion mid-keystroke.