Smart Combos
Set a model to auto or build your own. 17 routing strategies with
tier-1/2/3 fallback — quota out on one provider, switched to the next in milliseconds.
17 strategies
Free & Open-Source AI Gateway
Never Stop Coding.
236 providers, one endpoint, auto-fallback. Never hit limits — never stop building.
236 providers 90+ free 15–95% saved
$ omniroute
▸ cc/claude ........ ✓
▸ quota out → glm $0.5
▸ switched in 8ms
IDERouter236 providers
One endpoint. Never stop building.
-
Never hit limits
Auto-fallback across 236 providers in milliseconds. Quota out? The next provider takes over — zero downtime.
-
Save up to 95% tokens
RTK + Caveman stacked compression cuts 15–95% of eligible tokens — about 89% on tool-heavy sessions.
-
$0 to start
90+ providers with a free tier, 11 free forever (Kiro, Pollinations, LongCat…). No credit card needed.
-
Every tool works
16+ coding agents — Claude Code, Codex, Cursor, Cline, Copilot, Antigravity — through one config.
-
One endpoint
OpenAI ↔ Claude ↔ Gemini ↔ Responses API translation. Point any tool at
/v1and it just works. -
Production-grade
Circuit breakers, TLS stealth, MCP (95 tools), A2A, memory, guardrails, evals — backed by 21,000+ tests.
The catalog
providers
Most complete catalog of any open router.
90+ free 11 free forever
- 20 OAuth Claude Code, Codex, Cursor, Gemini CLI, Antigravity…
- 158 API-key OpenAI, Groq, NVIDIA, Cerebras, Mistral…
- 11 Free forever Kiro, Qoder, Pollinations, LongCat…
- 11 Local Ollama, LM Studio, vLLM…
23 Web 11 Search 7 Audio 24 Image 14 Video 14 Embeddings 8 Rerank 6 Music
Explore all 236 →Three commands. Zero config.
-
Install & run
One npm install, then launch — the API and dashboard come up together on port 20128.
-
Connect a FREE provider
Open the dashboard, pick from 90+ free tiers and sign in — no credit card, no paid API key needed.
-
Point your IDE
Set your tool's base URL to
localhost:20128/v1and your dashboard key. Every model is now one endpoint away. - Start Free →
$ npm install -g omniroute
added 1 package in 6s
$ omniroute
▸ dashboard ✓ http://localhost:20128/dashboard
▸ api ......✓ serving on :20128
$ curl localhost:20128/v1/models
✓ models listed 🎉
Everything built in. No add-ons.
Routing, resilience, free-quota draining, compression and agent protocols — one binary, zero plugins to glue together.
Memory
Persistent conversational memory — FTS5 keyword + Qdrant vector recall, built in.
hybrid recall
3-Layer Resilience
Fail at the right level — never the whole thing. Circuit breaker per provider, cooldown per connection, lockout per model. Other keys and models keep serving.
3 layers
Free Quota
Pool-deduped, honest counting across 90+ free providers.
~1.6B tokens / month
Compression
RTK + Caveman stacked on tool output and context.
15–95% tokens saved
MCP Server
Built-in server exposing the gateway itself — across 30 scopes.
95 tools / 30 scopes
A2A Protocol
JSON-RPC agent protocol — connect an agent, it drives OmniRoute.
6 agent skills
One config. Every coding tool works.
Point any OpenAI-compatible CLI at a single base URL — 236 providers behind the scenes.
- Base URL
localhost:20128/v1- API Key
‹from dashboard›
One endpoint. Every tool points here.
Claude Code
Codex
Cursor
Cline
Copilot
Gemini CLI
OpenCode
Kilo Code
Droid
Continue
Roo Code
Antigravity
Why OmniRoute
Every coding-AI pain has a built-in fix. No scripts, no babysitting.
❌ The daily pain
✅ How OmniRoute fixes it
Quota expires unused
Track & drain every free token
Rate limits stop you mid-task
4-tier auto-fallback in milliseconds
Tool output burns tokens
RTK + Caveman compression, 15–95%
Expensive APIs at $20–50/mo
Cost-optimized routing to cheapest
Each tool its own setup
One endpoint, every tool works
AI blocked in your region
3-level proxy + TLS stealth
IDE
① subscription
② API key
③ cheapest
④ free
How we compare
Audited against the real READMEs of 9router, LiteLLM and CLIProxyAPI. Where they win, we say so — where we're alone, the numbers speak.
| Feature | OmniRoute | 9router | LiteLLM | CLIProxyAPI |
|---|---|---|---|---|
| Routing | ||||
| Providers | 236 | 40+ | 100+ | 8+ upstreams |
| Routing strategies | 17 | 3-tier | retry / priority | round-robin · fill-first |
| Tier 1/2/3 fallback + UI | ✅ | ✅ | manual | ⚠ multi-account |
| Token compression | RTK+Caveman 15–95% | RTK 20–40% | ❌ | ❌ HTTP brotli only |
| Protocols | ||||
| Built-in MCP server | ✅ 95 tools — exposes the gateway | ❌ | ⚠ client (connects to MCP) | ❌ |
| A2A protocol | ✅ server, 6 skills | ❌ | ⚠ client (invokes A2A) | ❌ |
| Cloud agents (Codex/Cursor/Devin/Jules) | ✅ | ❌ | ❌ | ⚠ sidecar |
| Resilience & Security | ||||
| Resilience layers | 3 (breaker+cooldown+lockout) | ⚠ cooldown | ⚠ cooldown | ⚠ cooldown |
| Memory (FTS5 + vector) | ✅ | ❌ | ⚠ semantic cache | ❌ |
| Guardrails (PII / injection / vision) | ✅ | ❌ | ✅ | ⚠ cloak |
| Eval framework | ✅ | ❌ | ❌ | ❌ |
| TLS fingerprint stealth | ✅ wreq-js | ❌ | ❌ | ✅ utls |
| Ops | ||||
| Dashboard | Next.js 16 | Next.js 16 | Next.js | Next.js / React |
| i18n locales | 42 | 4 | ❌ | 3 |
| OAuth providers | 16+ | 5 | SSO | 6 (CLI auth) |
| Self-hostable | ✅ | ✅ | ✅ | ✅ |
| License | MIT | MIT | MIT + Comm | MIT |
| Stack | TS / Next 16 | Node / Next 16 | Python | Go |
Honesty notes: LiteLLM is an MCP/A2A client (connects/invokes); OmniRoute is a server (exposes its own gateway as 95 tools / 6 skills). CLIProxyAPI also ships TLS stealth (utls). 9router has RTK compression (20–40%) — the stacked RTK+Caveman (15–95%) is ours.
Where it runs
One codebase, eight ways to run it — laptop, server, phone or plugin. Pick the install command and you're serving on port 20128.
-
npm
npm i -g omniroute -
Docker
docker run omniroute -
Desktop
Electron app · Win/Mac/Linux -
ARM
arm64 · Raspberry Pi ready -
Termux
pkg install nodejs · run on Android -
PWA
Install from browser · offline-ready -
OpenCode plugin
@omniroute/opencode-plugin -
Source
git clone · npm run dev
Set your model to auto.
One alias, the whole catalog behind it. Pick a ready-made mode or build your own combo — 17 strategies, grouped by what you're optimizing for.
auto
/coding
/fast
/cheap
/offline
/smart
Or build your own — 17 strategies, grouped by goal:
-
🪫 Drain subscription
Burn the quota you already paid for before touching paid APIs.
priorityfill-first -
⚖️ Spread load
Balance traffic across keys and accounts to dodge rate limits.
round-robinp2cleast-used -
💸 Cheapest first
Always route to the lowest cost-per-token that can serve the request.
cost-optimized -
🧵 Context-aware
Keep long contexts on models that fit — relay when they overflow.
context-relaycontext-optimized -
🎲 Randomized
Unpredictable pick to avoid fingerprinting a single backend.
randomstrict-random -
🧠 Smart & adaptive
Let OmniRoute decide — 9-factor scoring, last-known-good, reset-aware windows.
autolkgpreset-awarereset-window
17 routing strategies 9-factor auto-scoring tier 1/2/3 fallback in milliseconds
Free tokens, honestly counted
~1.6B free tokens a month.
OmniRoute tracks every free tier across 90+ providers and drains the quota you'd otherwise lose. Usage is pool-deduped — shared accounts count once, so the numbers are real, not inflated.
90+ free providers 11 free forever
- Kiro 50 cr
- Qoder ∞
- LongCat 50M/day
- Cerebras 1M/day
- NVIDIA 40 rpm
- Pollinations no key
Pool-deduped & honest — no inflated multi-B claims.
Free forever — $0, no card:
Kiro
Pollinations
LongCat
Cloudflare- Gemini CLI
NVIDIA
Cerebras- + 4 more
Save 15–95% on tokens.
Tool output, repeated context and bloated JSON get compressed before they reach the model — same meaning, a fraction of the tokens. 10 composable engines stack in a pipeline you control.
The function should iterate over the list of items, and for each item, it should check whether the value is greater than zero, and if so, add it to the running total.sum all list items where value > 0Stacked in order — mix & match per combo:
- Session-Dedup drops content repeated across turns
- CCR archives big blocks, retrieves on demand
- RTK smart tool-result filtering & dedup
- Headroom lossless tabular compaction of JSON arrays
- Relevance extractive sentence scoring vs your last query
- Caveman rule-based prose compression — up to ~75%
- LLMLingua-2 ML semantic pruning (ONNX), code-safe
- Lite15%
- Standard30%
- Aggressive50%
- Ultra75%
- RTK60–90%
- Stacked78–95%
Output Styles — shape what comes back (combinable · lite/full/ultra):
- Terse prose
- Less code (YAGNI)
- Terse CJK 文言
All of it, under your control — header › combo › profile › adaptive › default:
- Named profiles + active selector
- Visual Compression Studio
- Anthropic Context Editing
- Adaptive context-budget dial
- Per-request header (x-omniroute-compression)
- Offline eval harness
- Auto-trigger by token threshold
Fail at the right level — not the whole thing.
A bad key shouldn't kill a provider. A dead model shouldn't kill a key. Each failure is isolated to the smallest possible scope, so everything else keeps serving.
Provider down? The breaker opens — combo routing skips it until it recovers.
Bad key? Cool it down — every other key for the same provider keeps serving.
One model quota-limited? Lock just that model — the rest of the models keep serving.
model ⊂ connection ⊂ provider — lazy recovery, no background timers. Our 3-layer model goes beyond the flat 3-tier fallback everyone else ships.
AI blocked in your country? Not anymore.
- 🇷🇺
- 🇨🇳
- 🇮🇷
- 🇨🇺
- 🇹🇷
- — reach all 236 providers
3-level proxy
Route through a proxy globally, per provider, or per individual connection.
- Global · per-provider · per-connection scope
- SOCKS5 & HTTP(S) proxies
- 1proxy marketplace, one-click
Stealth
Look exactly like the real CLI — down to the TLS handshake — so upstreams never flag you.
- TLS JA3/JA4 fingerprint via wreq-js
- CLI fingerprint match per provider
- IP preservation across requests
- IPv4 / IPv6 egress control
Private & local-first
Your keys, usage and history live on your machine — not someone else's cloud.
- SQLite on your disk · AES-256 encrypted credentials
- No telemetry, no account required
- Self-host anywhere — your data never leaves
A full dashboard, not a config file.
Providers, combos, analytics, health and costs — all on one screen at
localhost:20128/dashboard. No YAML editing required.
236
providers connected
90+
free tiers active
198
healthy now
- Claude (OAuth) live
- Cerebras live
- Groq limited
- NVIDIA NIM live
17
strategies
auto
default combo
8ms
avg failover
- auto on
- /cheap on
- /coding on
1.42M
requests · 30d
99.4%
success rate
312ms
p50 latency
3
resilience layers
2
on cooldown
0
breakers open
- Circuit breaker · provider CLOSED
- Cooldown · connection WAIT
- Lockout · model OK
$0.00
spent · 30d
1.4B
free tokens used
~$340
est. saved
- Free tiers
- Subscriptions
- Paid APIs
Connect an agent — it drives OmniRoute itself.
OmniRoute isn't just a proxy — it speaks the agent protocols. Plug it into an MCP client, an A2A network or a cloud coding agent and it becomes a tool the agent can operate.
-
MCP Server
Exposes the whole gateway as 95 tools across 30 scopes — routing, quota, memory, skills and more — over stdio, HTTP and SSE transports.
95 tools · 30 scopes · stdio / HTTP / SSE -
A2A Server
A full Agent-to-Agent server with 6 skills — smart routing, quota, discovery, cost analysis and health reporting — over JSON-RPC 2.0.
6 skills · JSON-RPC 2.0 · Agent Card -
Cloud Agents
Drive Codex, Cursor, Devin and Jules through one interface — create tasks, approve plans and stream results, all routed through OmniRoute.
Codex · Cursor · Devin · Jules
$ claude mcp add omniroute --type http --url http://localhost:20128/api/mcp/stream
✓ omniroute connected — 95 tools available
Not just a proxy — a platform.
-
Embedded services
Run Bifrost (Go relay), 9Router or CLIProxy as managed sidecars — or a full cluster profile — straight from the dashboard.
-
Your knowledge, in context
Notion & Obsidian wired in as MCP tools — your notes and vaults become first-class context for any model.
-
Guardrails built in
Prompt-injection guard on every route, opt-in PII redaction and safety content-filter normalization — defense without lock-in.
-
Gamification
Streaks, achievements and live savings — your usage, made tangible.
Watch OmniRoute run — in your language
Full walkthroughs: install, connect a free provider, point your IDE — and never hit a limit again.
A fast-moving, active community
OmniRoute ships constantly — driven by hundreds of contributors who keep it ahead of the curve.
Building in the open
since Feb 13, 2026 122 days non-stop
- 8,129GitHub stars
- 1,367forks
- 270+contributors
- 271releases shipped
- 2,969PRs merged
- 1,746issues closed
- 21,000+automated tests — quality first