OmniRoute

An MIT-licensed, self-hosted AI gateway by diego-souza: one local OpenAI-compatible endpoint (http://localhost:20128/v1) fronting a claimed 290+ providers and 500+ models, of which 90+ have free tiers. The README tagline is “Never stop coding. Free MIT AI gateway: one endpoint, 290+ providers (90+ free), 500+ models.”

T1 by origin (the project’s own repo) but every number below is self-reported and unaudited — provider counts, token budgets and compression ratios are the vendor’s, not a third party’s.

What it actually does

• Cascade fallback in four tiers — subscription seats (Claude, Copilot) → API-key providers (deepseek, Groq) → cheap providers (GLM, MiniMax) → free providers (Kilo, OpenCode, Pollinations). When a tier’s quota dies the request drops to the next one instead of failing.
• 19 routing strategies, named: priority, fill-first, weighted, round-robin, p2c, least-used, random, strict-random, cost-optimized, headroom, reset-window, reset-aware, context-relay, context-optimized, cache-optimized, lkgp, auto, fusion, pipeline. model: auto picks for you.
• A 12-engine compression pipeline (Session-Dedup, CCR, Lite, RTK, Responses Tool Output, Headroom, Relevance, Caveman, Aggressive, LLMLingua-2, Ultra, OmniGlyph) claiming 15–95% token savings, 89% “average” — a range, not a guarantee, and the README says so.
• Three independent resilience layers: provider circuit breakers, connection cooldowns, model-level lockouts.
• Local-first: runs on your machine, credentials AES-256-GCM encrypted, SQLite state, a Next.js dashboard for quota/cost, 104 MCP tools so an agent can drive the gateway itself.

The free-tier ledger

The headline economic claim: ~1.53B free tokens/month aggregated across 43 provider pools and 516 models, rising to ~2.15B in the first month with signup credits. Named free-forever pools include OpenCode Zen (no token cap), Kilo Code, Requesty, SiliconFlow, Z.AI GLM, Baidu ERNIE, Qoder AI, Pollinations, Cloudflare AI, NVIDIA NIM and Cerebras. The project re-audits the figures every two weeks and notes they “move both ways.” It also flags 15 providers as ToS-questionable and leaves the call to the user. Some of that 1.53B is capacity the provider never meant to give away, and the project says so itself.

Where it sits

v3.8.49, ~30.4K stars, 4K forks, 5,749 commits, biweekly releases; TypeScript/Node, pnpm monorepo, shipped as npm, Docker, Electron, PWA and Termux/Android. The README carries a comparison table against 9router, OpenRouter, CLIProxyAPI and LiteLLM across 13 capabilities, claiming to be the only one with the complete set — a self-scored table, so read it as positioning.

Related

ai-gateway · llm-api-pricing · amazon-bedrock · llm-provider · diego-souza