Features
No marketing language. Every feature below is a specific capability with audited tests, listed by what it changes for you day-to-day.
Cost
The cheapest call is the one you never make. RTK shaves the bloat from tool outputs before they reach the model. Caveman trims the response. The dashboard shows the dollars you didn't spend.
Detects tool_result content (git diff, grep, ls, tree, log dumps) and applies lossless compression before the request leaves your machine. Auto-detects type, fails open safely, runs before format translation so it works across every client. Default ON.
Optional terse-style system prompt injection. Models reply with technical substance, dropped articles and filler. Saves up to 65% output tokens on factual queries. Off by default.
Dashboard compares what you actually spent against what paid APIs would have charged. See the dollar value of kRouter in real time.
Routing
Quota walls don't interrupt you. Combos chain subscription, cheap, and free providers into a single endpoint. When one tier exhausts, the next picks up. Your IDE never sees an error.
Kodelyth's proprietary AI-driven scoring engine replaces dumb failovers. Zenith evaluates live health data (TTFB latency, success rate) and quota headroom, mathematically pre-ranking accounts to eliminate rate-limit stalls.
An in-memory HealthCache completely eliminates SQLite reads from the hot path. When an account hits a 429, the router instantly locks it in RAM and grabs the next best account in < 1ms.
Subscription → cheap → free. Combos compose freely and can reference other combos.
Claude Code, Codex, GitHub Copilot, Cursor, Antigravity, Kiro, OpenRouter, NVIDIA, Vertex, GLM, MiniMax, Kimi, Groq, xAI, Mistral, Perplexity, Together, Fireworks, Cerebras, Cohere, DeepSeek, and more.
OpenAI ↔ Claude ↔ Gemini ↔ Cursor ↔ Kiro ↔ Vertex ↔ Antigravity. Your tool sends one format, kRouter rewrites to native upstream shape, response comes back unified.
Tokens refresh before expiration. Concurrent-safe — no stale-401 cascades. Atomic DB writes. You never re-login.
Observability
kRouter is a glass-walled router. Every account, every model, every request — live numbers, no black boxes.
Per-model token counts pulled live from each provider. Reset countdowns (5h / daily / weekly). Distinguishes 0% remaining from 'no fraction reported'. TPM vs daily-quota disambiguation on 429.
Toggle debug mode for full request/response. Headers, payloads, latency, account picked, fallback chain — all visible in dashboard. Export as JSONL.
Sync providers, combos, and settings across devices via encrypted background sync. Off by default. Fail-fast on cloud unreachable so the UI never hangs.
Security
kRouter runs on your machine. Credentials never leave SQLite. Every request is observable, every error is honest, every cert is your own.
Intercept Antigravity, Kiro, Copilot, and Cursor IDE traffic locally. Bounds-checked encoding (name ≤ 255, value ≤ 65535, frame ≤ 16 MiB). NGHTTP2 stream-recovery → HTTP/1.1 fallback. Parseable Kiro exception frames on upstream errors.
Expose the dashboard over a public tunnel or Tailscale. Configurable access controls block dashboard login via tunnel unless explicitly enabled.
Token refresh no longer mutates caller credentials under concurrency. Atomic backoff increments via SQLite transaction. Timing-safe CLI token compare. Per-IP brute-force lockout.
User-supplied baseUrl is validated: blocks cloud-metadata endpoints (AWS / GCP / ECS / Alibaba) and non-http(s) schemes. Loopback and private LAN ranges remain allowed for self-hosted models.
Ready when you are
Install in 30 seconds. Connect a provider, point your IDE, keep coding. MIT licensed. No card. No vendor lock-in.