Local LLM proxy that saves you tokens. Runs as a background daemon, intercepts requests from AI agents (Claude Code, Cursor, OpenClaw), and routes them to the cheapest suitable provider — automatically.
Your AI agent thinks it's talking to Anthropic / OpenAI / etc. Behind the scenes, krouter:
- Routes to the cheapest provider that meets your quality bar
- Tracks cost per request, per model, per agent
- Honors your existing API keys (they never leave your machine)
- Saves you tokens — often 50-80% with the same outcomes
┌──────────────────────┐ ┌──────────────────────┐
│ Claude Code / │ localhost │ krouter │
│ Cursor / OpenClaw │ ───────────────► │ daemon (8402) │
└──────────────────────┘ └──────────┬───────────┘
│
┌──────────────────────────┼──────────────────────────┐
↓ ↓ ↓
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ Anthropic │ │ DeepSeek │ │ Moonshot │
└──────────────┘ └──────────────┘ └──────────────┘
(router decides at each request)
Live routing — see exactly what each agent asked for and where krouter actually sent it, with the savings per request:
Connected apps — krouter auto-detects your AI agents and inherits their provider configs; no keys re-entered:
- Forever free, MIT licensed — no paid tier, no subscription
- Zero data upload — no telemetry, no backend dependency
- API keys never persisted — they pass through, never touch disk
- Zero-elevation install — no
sudo, no admin password - Cross-platform — macOS, Linux, Windows
- Same-protocol routing — Anthropic requests stay Anthropic-protocol
| Provider | Env var | Models |
|---|---|---|
| Anthropic | ANTHROPIC_API_KEY |
claude-3-5-sonnet, claude-3-haiku, ... |
| OpenAI | OPENAI_API_KEY |
gpt-4o, gpt-4o-mini, ... |
| DeepSeek | DEEPSEEK_API_KEY |
deepseek-chat, deepseek-reasoner |
| Groq | GROQ_API_KEY |
llama-3.3-70b, mixtral-8x7b, ... |
| Moonshot | MOONSHOT_API_KEY |
moonshot-v1-8k/32k/128k |
| GLM (Zhipu) | ZHIPU_API_KEY |
glm-4, glm-4-flash, ... |
| Qwen (Alibaba) | DASHSCOPE_API_KEY |
qwen-max, qwen-plus, qwen-turbo |
Pre-release builds are in development. Release packages will be available in Releases.
1. Download krouter-X.X.X-macOS.dmg from Releases
2. Open the .dmg and drag to Applications
3. Launch — the install wizard handles the rest
1. Download krouter-X.X.X-setup.exe from Releases
2. Run the installer (no admin required)
# Debian/Ubuntu
sudo dpkg -i krouter_X.X.X_amd64.deb
# RPM-based
sudo rpm -i krouter-X.X.X.x86_64.rpm
After install, the wizard auto-detects and connects your AI agents. Then just use them as usual.
Set your provider API keys in your shell environment (or the GUI settings panel):
export ANTHROPIC_API_KEY=sk-ant-...
export DEEPSEEK_API_KEY=sk-...krouter status
krouter logs --follow
krouter budget
krouter config set preset saver
krouter test # smoke-test routingRequires Go 1.22+, CGO enabled.
# Build daemon + CLI
make build
# Run tests
make test
# Lint
make lint
# Run in dev mode
make devTo build the install wizard:
make build-installer # Go binary
make build-frontend # React frontendsThe daemon:
- Sends requests only to LLM providers you configured (your keys, your providers)
- Polls
announcements.kinthai.ai/feed.jsonevery 6 hours (anonymous GET, no identifiers) - Checks GitHub Releases every 24 hours for updates (anonymous GET)
We never:
- Upload any data to any backend
- Track user identifiers, emails, or usage patterns
- Store your API keys (they're forwarded from your env at request time)
Three processes:
krouter(daemon) — runs as LaunchAgent / systemd --user / Windows Task Schedulerkrouter-installer— one-shot install wizard, serves browser UI at :8404krouterweb UI — embedded React app served by the daemon at :8403/ui/
Three ports:
- 8402 (proxy) — agent-facing, always 127.0.0.1, no auth
- 8403 (management) — web UI + API, 127.0.0.1 by default, optional 0.0.0.0 for LAN remote
- 8404 (installer) — install wizard, only active while
krouter-installerruns
PRs welcome. Please open an issue first for significant changes.
MIT — see LICENSE.
- OpenClaw — managed OpenClaw fork
- kinthai.ai — agent marketplace