RTK Token Saver
Compresses tool_result content (git diff, grep, ls, tree) inline before sending. Saves 20–40% input tokens per request.
One endpoint.
Build Everything.
The AI infrastructure layer beneath your favourite IDE. Route Claude Code, Cursor, and 98+ providers through a single OpenAI-compatible endpoint on your machine.
- MIT licensed
- Self-hosted
- 12+ IDEs supported
- 98+ providers
- 3.3k+ monthly downloads
localhost:20128
v1 chat · completions
01 · Your IDE
cc/
cur/
ag/
gh/
+802 · kRouter
kRouter:20128
~5s
latency
40%
savings
$0
cost
03 · Providers
Claude
OAuth · Pro/Max
Codex
OAuth · Plus/Pro
Vertex
$300 credits
Kiro
Free · unlimited
Catalog+ 94 more
OpenAI-compatibleprompts stay localMIT
How it works
One local endpoint.
Every provider behind it.
kRouter runs on your machine and exposes a single OpenAI-compatible endpoint at localhost:20128. Point any IDE at it. Behind the scenes it routes to your subscription first, falls back to free tiers, and only touches paid keys when nothing else works.
- No proxy in the cloud. Your prompts stay on your machine.
- Quota-aware. Switches providers automatically when one runs out.
- Format translation. OpenAI ↔ Claude ↔ Gemini — any tool, any model.
Your IDE
- Claude Code
- Cursor
- Antigravity
- Copilot
- Codex
- Kiro
kRouter
:20128
98+ providers
- Anthropic
- OpenAI
- GitHub
- Kiro
- Mistral
- Groq
- +91 more
one OpenAI-compatible API · subscription → free → paid fallback · RTK compression · format translation
Install in 30 seconds
One command. Three ways.
Pick NPM if you just want it to work. Docker if you self-host. Git if you want to hack on the code.
$
npm install -g @sifxprime/krouter$
krouter -tDashboard: http://localhost:20128/dashboard
Built for production
Real features from the repo — not marketing slides.
Every feature ships with audited unit tests, end-to-end verification, and visible-in-the-dashboard observability.
Caveman Mode
Optional terse-response prompt injection (Lite / Full / Wenyan) to cut output tokens up to 65% without losing technical substance.
3-Tier Fallback
Subscription → cheap → free. Combos auto-rotate when one quota hits zero. No more rate-limit interruptions.
Routing Strategies
Per-provider account selection: fill-first, round-robin (sticky), p2c, or random. Atomic backoff keeps concurrency safe.
Response Cache
In-memory cache for repeated non-streaming requests. Warmup probes and title generation skip redundant upstream calls.
Real-Time Quota
Live token counts, reset countdowns, and per-model spend. Quota tracker matches Google's own backend numbers.
MITM Mode
Intercept Antigravity, Kiro, Copilot, and Cursor IDE traffic locally. Bounds-checked frames, NGHTTP2 stream recovery.
Tunnel / Tailscale
Expose the dashboard over a public tunnel or Tailscale. Configurable access controls so it stays safe when shared.
Provider Ecosystem
OAuth, free credits, API key, and browser-cookie providers. Stack them into combos that auto-fall-through.
Format Translation
OpenAI ↔ Claude ↔ Gemini ↔ Cursor ↔ Kiro. Use any tool with any model. The translation layer handles the rest.
Auto Token Refresh
OAuth tokens refresh before expiration. Concurrent-safe — no stale-401 cascades when traffic spikes.
Zero Cost
MIT licensed, self-hosted, never charges. The dashboard 'cost' is a savings tracker — you only pay providers directly.
Works with your stack98+ providers.
Browse all providers98+ providers.
One endpoint.
OAuth subscription providers, genuinely-free tiers, and pay-per-token APIs. Stack them into combos so requests auto-fall-through.
Kiro AI
kr/
OpenCode Free
oc/
OpenRouter
openrouter/
NVIDIA NIM
nvidia/
Vertex AI
vx/
Gemini
gemini/
BytePlus ModelArk
bpm/
Claude Code
cc/
Antigravity
ag/
OpenAI Codex
cx/
GitHub Copilot
gh/
Cursor IDE
cu/
xAI (Grok)
xai/
GLM Coding (Z.ai)
glm/
Kimi
kimi/
Minimax Coding
minimax/
OpenAI
openai/
Anthropic
anthropic/
DeepSeek
ds/
Groq
groq/
Mistral
mistral/
Perplexity
pplx/
Together AI
together/
Fireworks AI
fireworks/
How we compare
See full 5-way comparisonHonest, side-by-side.
All three projects share a common ancestor (CLIProxyAPI in Go). Each takes the idea in a different direction. Pick by fit, not loyalty.
FAQ
Questions people actually ask
Is kRouter free?
Yes. kRouter is MIT-licensed, open-source, and self-hosted — it never charges you anything. You only pay the underlying AI providers you connect, and many of them (Kiro, Antigravity credits, OpenCode, Gemini free tier, and more) have generous free tiers. The gateway itself is free forever.
What is kRouter, and how is it different from an LLM router?
kRouter is a self-hosted multimodal AI gateway. It exposes one local OpenAI-compatible endpoint (http://localhost:20128) that routes eight modalities — chat, text-to-speech, speech-to-text, image, video, embeddings, web search, and web fetch — across 98+ providers. A plain "LLM router" only switches between chat models; kRouter unifies every modality behind a single base URL and SDK.
How do I use Claude Code without paying $200/month?
Point Claude Code (or Cursor, Cline, Aider) at kRouter's local endpoint, then connect a free Claude backend like Kiro. kRouter's 3-tier fallback runs your free tier first, a cheap pay-per-token model as overflow, and only touches a paid key when nothing else is available — so most requests cost nothing.
How much does kRouter save on tokens?
Up to 92%. kRouter ships four token savers: RTK (lossless input compression, 20–40%), Caveman Mode (terse output, up to 65%), Headroom (context de-duplication, measured 92.4% on code-heavy requests), and PXPIPE (context-to-image compression). They stack — running RTK + Headroom + Caveman together on a long agentic session cuts token cost dramatically.
Does kRouter keep my prompts private?
Yes. kRouter runs entirely on your own machine (or your own server). Your prompts, code, audio, and documents go straight from your machine to the provider you configured — there is no third-party proxy in the middle logging your data. It is self-hosted by design.
Which IDEs and tools work with kRouter?
Any OpenAI-compatible client: Claude Code, Cursor, Cline, Aider, Continue, Roo, Windsurf, Kiro, OpenCode, Kimi, Gemini CLI, and more. kRouter also intercepts IDE traffic via MITM mode for clients that hardcode their endpoint, and translates between OpenAI, Claude, Gemini, Vertex, and Ollama request formats automatically.
Can I run kRouter on a server for my whole team?
Yes. Deploy kRouter on a $5/month VPS behind nginx or Tailscale and point every teammate's IDE at the shared endpoint so quota pools across the team. Headless providers like Kiro now support API-key (ksk_) auth, so you can connect them without a browser login on a server.
How do I install kRouter?
Run "npm install -g @sifxprime/krouter && krouter -t", then open the dashboard at http://localhost:20128/dashboard, connect a provider, and point your IDE at http://localhost:20128. Docker and source-build installs are also supported. It takes about 30 seconds.
The new floor for AI coding
Stop choosing tools.
Build Everything.
kRouter is the layer beneath your favourite IDE — Claude Code, Cursor, Antigravity, Copilot, Codex, Kiro. Connect any model. Switch providers on quota. Keep every prompt on your machine.
MIT licensed · v0.5.118 · Built in the open.
What you actually get
- Local endpointOpenAI-compatible API at localhost:20128
- Providers98+ — OAuth, free, and pay-per-token
- ModalitiesChat · TTS · STT · Image · Video · Embeddings · Web search
- IDEsClaude Code · Cursor · Antigravity · Copilot · Codex · Kiro · Cline · Roo · Aider · OpenCode · Kimi · Gemini CLI
- Token savingsUp to 92% via RTK · Caveman · Headroom
- PrivacyPrompts never leave your machine
- LicenseMIT — self-hosted forever