Open Source
Open Source

OmniRoute: One Endpoint to 290 Models, Letting Claude Code Run on Free Claude and GPT

This week's GitHub weekly #1 (+11,190 stars/week) is a free AI gateway: OmniRoute aggregates 290 providers (90+ free) into one local endpoint, with 33 coding agents (Claude Code/Cursor/Cline) zero-config; 4-tier auto-fallback + 12-engine compression saving 89% tokens, ~1.53B free tokens/month. Three-minute setup, pitfalls and compliance risks.

Published July 27, 202610 min read
<!-- omniroute-free-ai-gateway-resource | resource | OmniRoute Free AI Gateway -->

This week's GitHub weekly ranking has a runaway champion: OmniRoute, which pulled 11,190 stars in seven days, nearly 4,000 ahead of second place. 29.9k total stars, MIT license, 500+ contributors, currently v3.8.49. What it does fits one sentence. It spins up a local port on your machine, turns 290 AI model providers (90+ with a free tier) into one endpoint, and lets coding agents like Claude Code, Cursor, and Cline run on free Claude, GPT, and Gemini. Maintainer Diego is Brazilian, and this is the first Brazilian open-source project in Kimi's support program.

What Pain It Solves

Anyone using a coding agent has been ground down by these: an API key suddenly rate-limits mid-edit; your Claude Code subscription quota resets unused at month-end; tool output (git log, build logs) burns tens of thousands of tokens; every tool needs its own key setup, a dozen dashboards to juggle; prompts all run through someone else's cloud, so sensitive code never feels safe; a team sharing one subscription ends up in a grab-what-you-can race. OmniRoute bundles the fixes: hit a limit and it auto-switches to the next provider, subscription quota gets drained first, tool output runs through a compression pipeline, all keys are encrypted locally with AES-256-GCM, data never leaves your machine, and teams can share a key pool with fair quotas that lend idle slices out when nobody's using them. The README lists ten daily pains mapped to ten fixes. The core message: stop juggling a dozen dashboards and dead keys.

One Endpoint, 33 Tools

After install you hand the client one address: http://localhost:20128/v1. It speaks the OpenAI-compatible protocol, so anything that eats an OpenAI interface works. The official list has 33: Claude Code, Codex, Cursor, Cline, Copilot CLI, Aider, Continue, Goose, Open Interpreter. Routing falls back across four tiers: your subscriptions first (Claude Code/Codex/Copilot), then your own API keys (DeepSeek/Groq/xAI), then a cheap tier (GLM $0.5, MiniMax $0.2), finally a permanent free tier (Kiro, Qoder, Pollinations). Switching is millisecond-level and invisible to the client. Three self-healing layers back it: a provider-level circuit breaker (trips only on 5xx, half-open probe after 60s), a key-level cooldown with backoff (429 honors Retry-After; one cooling key never blocks siblings), and a model-level lockout (one model dying never takes down the whole connection).

19 Routing Strategies Plus Auto

Don't want to configure? Set auto. It scores on 12 factors (health, quota, cost, latency, success rate) to pick the best provider right now. Auto splits into variants: auto/coding weighs quality for code, auto/fast chases latency, auto/cheap chases cheap, auto/offline chases quota headroom, auto/smart leaves 10% to explore better models. To control it yourself, 19 strategies mix and match: priority drains an ordered list, fill-first squeezes each quota dry, cost-optimized minimizes dollars, context-relay hands long conversations across models, fusion fans out to a panel then has a judge synthesize one answer, pipeline chains targets into a sequence, lkgp sticks to the last successful provider and doesn't bounce, cache-optimized pins identical prompt prefixes to the same account to maximize cache hits. A combo is a chain across models; when quota runs out or a provider fails, it silently slides to the next. That's the "unbreakable" claim.

How It Saves 89% of Tokens

The counterintuitive part is compression. Default runs RTK plus Caveman stacked, tool and context payload pass through both, saving 78-95%, averaging 89.2% on tool-heavy sessions. The official example: a 69-token React explanation compressed to 19 tokens, same answer, 72% saved. Compressed it reads: "New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo." The key point: code blocks, URLs, and JSON are always preserved byte-perfect and never compressed. Twelve engines compose: Session-Dedup drops content repeated across turns, CCR archives large blocks behind on-demand markers, RTK filters command output in tool results, Relevance does extractive sentence scoring against the last query, Caveman rule-compresses prose, LLMLingua-2 does ML semantic pruning via MobileBERT (code-safe, async), OmniGlyph experimentally encodes context as an image routed to Claude Fable 5 (most aggressive, opt-in). Presets tier up: Lite trims whitespace for 15% as a baseline, Standard saves 30% for daily use, Aggressive saves 50% for long tool sessions, Ultra saves 75% at the limit. An adaptive dial replaces the single on/off threshold: it escalates only the cheapest, lossless engines just enough to fit the model's context window. Aggressive modes have an eval harness (npm run eval:compression) to test fidelity on a pinned corpus first. Zero client changes; compression happens transparently in the pipeline.

How Much Free Tier There Really Is

The official number: about 1.53 billion free tokens per month. It aggregates the free tiers of 43 provider pools and 516 models, deduped into an honest figure, not the inflated number you'd get counting every rate limit 24/7. Free forever, no token cap: OpenCode Zen (DeepSeek V4), Kilo Code (Tencent Hy3), SiliconFlow (DeepSeek V3.2/R1), Z.AI GLM, Baidu ERNIE, Qoder (Qwen3-Max, Kimi-K2), Pollinations, Cloudflare AI, NVIDIA NIM, Cerebras. The number is re-audited every two weeks against the live catalog; a provider kills a free tier and it drops, a new one lands and it climbs. More honest than "free forever" marketing. First month brings signup credits (vertex 300M, agentrouter 200M, together 25M and more), and a $10 OpenRouter top-up unlocks an extra 24M tokens/month at higher RPM. A /dashboard/free-tiers page shows used vs. remaining live, with a per-provider quota grid. So it's not an "unlimited free Claude" magic trick; it's a tool that stitches scattered free quotas into a visible budget sheet, and you graduate to paid keys when you need to.

Three-Minute Setup

bash
# 1. Install and run
npm install -g omniroute
omniroute
# dashboard at http://localhost:20128, API at /v1

# 2. dashboard -> Providers, connect a free one
#    OpenCode Free needs no login, or Kiro AI (~50 free Claude calls/month)

# 3. In your coding tool:
#    Base URL: http://localhost:20128/v1
#    API Key:  copy from dashboard -> Endpoints
#    Model:   auto

# 4. Verify
curl http://localhost:20128/v1/models -H "Authorization: Bearer YOUR_KEY"

Fresh install works with zero keys and zero config; auto already replies (OpenCode Free and Felo are pre-wired into the default combo). There's a Docker image, a desktop app, and even a Termux build to run it on your phone. If a client can't send custom headers, it exposes tokenized compatibility aliases (/vscode/YOUR_KEY/...) for tools that refuse to attach an Authorization header. A single setup-* command auto-configures 12+ coding tools, and omniroute launch is zero-config straight into a run.

The People and the Engineering Behind It

The project was started by Diego Souza, maintained openly on GitHub, with 500+ contributors, a UI in 43 languages, and a test suite north of 25,000 cases. It's listed on Trendshift, with communities on Discord, Telegram, and WhatsApp. The engineering trade-off worth noting: it stuffs compression, routing, circuit breaking, MCP (104 tools), A2A, persistent memory, and prompt-injection guardrails into one process, 100% MIT self-hosted, no cloud-hosted pay tier. That's the source of its "only full-featured set" claim in the comparison table; the cost is a dense config surface and a learning curve. It also ships a transparent MITM decrypt (TPROXY) to capture CLIs that ignore proxy env vars, and its memory is off by default, using int8 vector quantization with decay when enabled.

Who It's For, and Caveats

For: developers using coding agents daily who want to cut cost or resist rate limits; teams sharing subscriptions who need fair quota sharing; the privacy-sensitive who don't want prompts crossing a cloud.

Four caveats. First, compliance: 15 providers carry a ToS risk flag (using someone's free tier for commercial work is your call to make), shown in the dashboard. Second, aggressive compression occasionally drops detail in long sessions; don't start important tasks on Ultra. Third, some README links carry aff=omniroute sponsorship tags (no extra cost to you, but worth knowing). Fourth, the free tier is stitched from many providers' free tiers with rate limits, not unlimited; heavy loads still need a paid key. Fifth, on the engineering side: it installs as a global npm package, an old Node version throws ERESOLVE warnings (harmless per the docs), Windows has occasional path quirks, and Docker is the low-friction option. A remote mode runs OmniRoute on a VPS with the local CLI connecting in, suited for teams that need 24/7 uptime and shared access.

Versus Competitors, and the Kimi K3 Connection

Against OpenRouter and LiteLLM, OmniRoute's official comparison table (13 capabilities) shows it's the only one with the full set: 290 providers, 19 routing strategies, 12-engine compression, MCP 104 tools, A2A, persistent memory, TLS fingerprint stealth, all platforms, 100% MIT self-hosted. OpenRouter leans cloud aggregation, LiteLLM is a proxy library, different positions.

One more connection. OmniRoute's Founding Open Source Friend is Kimi (Moonshot AI). Its PR merge validation runs on Kimi K3, and both the Kimi API and Kimi Code subscription are first-class integrations. K3, whose weights opened on July 27, works out of the box in this gateway. One side builds the largest open model, the other wires open models into every coding tool. Upstream and downstream click together.


OmniRoute itself isn't complicated. What's complicated is operating the 290 providers it aggregates behind the scenes. Install it, set auto, and what you save is the time spent switching keys, checking quota, tallying cost. To freeload smartly, don't just set auto and walk away. Open the dashboard's free-tiers page and know whose quota you're burning. It's still #1 on the weekly chart this week, iterating fast; pin a version before you ship.


References

This article is AI-assisted and human-edited. Last updated: 2026-07-27

FAQ

What is OmniRoute? Is it really free?
A local AI gateway: npm install one, it runs an OpenAI-compatible endpoint at localhost:20128/v1, wiring in 290 providers (90+ with free tiers). Zero keys, zero config on install; set model to auto and it uses pre-wired free providers. The free tier aggregates many providers' free tiers to ~1.53B tokens/month, rate-limited, not unlimited.
Does 89% token compression hurt code quality?
Default RTK + Caveman stacked saves 78-95%. Code blocks, URLs, and JSON are always preserved byte-perfect, never compressed. Aggressive modes (Ultra, 75%) occasionally drop detail in long sessions; use Standard (30%) for important tasks or run the eval harness first.
Which models are free forever?
OpenCode Zen (DeepSeek V4), Kilo Code (Tencent Hy3), SiliconFlow (DeepSeek V3.2/R1), Z.AI GLM, Baidu ERNIE, Qoder (Qwen3-Max/Kimi-K2), Pollinations, Cloudflare AI, NVIDIA NIM, Cerebras and more, no token cap. The full list is re-audited every two weeks.
Who is it for? What are the risks?
For developers using coding agents daily who want to cut cost/resist rate limits, teams sharing subscriptions, privacy-sensitive setups. Caveats: 15 providers carry a ToS risk flag (commercial use is your call), aggressive compression may drop detail, some links carry aff sponsorship, and the free tier has rate limits so heavy loads need a paid key.

Related

Open Source

video-shotcraft: Turn Claude Code Into a Cinematic Product Video Studio (3.5K Stars)

video-shotcraft (Vincentwei1021/video-shotcraft, 3.5K GitHub stars, TypeScript, Apache-2.0) is an agent skill that turns Claude Code or Codex into a motion-design studio - point it at your product and it storyboards, animates, and sound-designs a cinematic promo with Remotion. Ships 104 shot recipe cards, 161 motion previews, a validated 36.2s Ink Press template, 2.5D camera moves, beat-synced cuts, and film-grade SFX. Output is reproducible TSX, not black-box generation. Includes a China AtomGit mirror and 3 headless-rendering CI flags.

Aug 4, 20268 min read
Open Source

book-to-skill: Turn Any Technical Book Into an AI Agent Skill (16.2K Stars)

book-to-skill (virgiliojr94/book-to-skill, 16.2K GitHub stars, Python, MIT) distills technical books/PDFs/EPUBs/doc folders into structured agent skills following the open Agent Skills standard - install once, works across GitHub Copilot CLI, Amp, and Claude Code. Generates SKILL.md + per-chapter files + glossary + patterns + cheatsheet, with chapters loaded on-demand so they don't count against your token budget. Ships a benchmark tool measuring 24-51x fewer tokens than dumping the full book into context (tested on 3 real books). Beyond books: internal docs, brand systems, research clusters, specs. Includes a copyright-compliance note.

Aug 4, 20267 min read