decolua/9router

A local OpenAI-compatible router that connects AI coding tools to many model providers with tiered auto-fallback and tool-output token compression.

What it is

9Router is a self-hosted gateway that sits between AI coding CLIs (such as Claude Code, Codex, Cursor, Cline, and others) and 40-plus model providers. It exposes an OpenAI-compatible endpoint on localhost, translates request formats between providers, tracks quota, and automatically falls back across a subscription → cheap → free tier so work continues when one provider hits a limit. It also compresses tool-result content before sending it to the model to reduce input tokens. The audience is developers who use multiple providers or subscriptions and want a single routing layer with a dashboard.

Key features

RTK Token Saver: auto-detects tool outputs (git diff, git status, grep, find, ls, tree, logs) and applies compression before the request reaches the LLM; it runs before format translation and silently keeps the original text if a filter fails or grows the output.
Three-tier fallback across subscription, cheap, and free providers, configured as named "combos" that switch automatically on quota exhaustion or errors.
Format translation between OpenAI, Claude, Gemini, Cursor, Kiro, Vertex, Antigravity, Ollama, and OpenAI Responses formats.
Multi-account support per provider with round-robin or priority routing, plus automatic OAuth token refresh.
Real-time quota tracking, usage analytics, request logging/debug mode, and cloud sync of config across devices.
Optional Headroom integration (an external /v1/compress proxy) and prompt-injection modes (Caveman, Ponytail) for reducing output tokens.
Multiple deployment targets: localhost, VPS, Docker, and Cloudflare Workers.

Tech stack

Primary language: JavaScript; the web dashboard is a Next.js application (next ^16.1.6) with React 19.2.4.
Server/runtime: express ^5.2.1, http-proxy-middleware ^3.0.5, undici ^7.19.2, jose ^6.1.3 for JWTs, node-forge and selfsigned for certificates, socks-proxy-agent.
Storage: better-sqlite3 ^12.6.2 kept as an optional dependency with sql.js ^1.14.1 as a runtime fallback for systems without build tools.
UI: @monaco-editor/react, @xyflow/react, recharts, @dnd-kit/*, material-symbols, zustand, Tailwind CSS v4.
Distributed as a global npm CLI (npm install -g 9router) and via Docker/GHCR images; a separate cli/ package builds the terminal UI and system tray.

When to reach for it

You use one or more AI coding CLIs and want a single local endpoint that routes across many providers.
You want automatic failover when a subscription or free tier runs out of quota mid-task.
You juggle multiple accounts or subscriptions and want round-robin load balancing and quota visibility.
You want to cut the token cost of large tool outputs without changing your client.

When not to reach for it

You use a single provider and model steadily and do not need routing or fallback.
You need a stable, versioned API — the app package is pre-1.0 (0.5.40) and carries a very large open-issue count.
You are uncomfortable routing provider credentials and OAuth tokens through a local proxy layer.
You are latency-sensitive, since the router adds a network hop plus format translation and compression on each request.

Maturity signal

9Router gained a large audience quickly, reaching nearly 23,000 stars within roughly half a year of its creation, and it is under very active development with frequent pushes. At the same time it is pre-1.0 and carries over a thousand open issues, which points to rapid, still-stabilizing development rather than a settled release. Treat it as fast-moving and popular but early: expect breaking changes and churn, and pin versions if you depend on it.

Alternatives

LiteLLM — use it when you want a widely adopted, provider-agnostic proxy focused on a unified API and routing rather than a coding-tool dashboard.
OpenRouter — reach for it when you prefer a hosted aggregation service over running a local router (9Router can itself route through it as one provider).
Portkey — consider it when you want a gateway emphasizing observability, guardrails, and caching for production API traffic.

Notes

The root npm package is private (9router-app), so the README treats running from source or Docker as the expected local development path, while the published 9router npm package is the CLI. The dashboard's displayed "cost" is explicitly a tracking and savings estimate, not a bill — 9Router itself never charges. The README also documents that several free provider tiers (iFlow, Qwen Code, Gemini CLI) were discontinued in 2026 and that Kiro AI moved to a paid model, so the free-tier landscape it targets shifts frequently.

Wiki: decolua/9router