What it does

9Router sits between your AI coding tools and the internet. You point Claude Code, Cursor, Cline, Copilot, or Codex at localhost:20128/v1 instead of the provider’s real endpoint. It routes requests through paid subscriptions first, then cheap APIs, then free tiers when quotas run dry. It also claims to compress tool outputs (git diffs, grep results) to shave 20–40% off token counts.

The interesting bit

The “RTK Token Saver” is doing the boring work that actually matters: stripping and compressing tool_result blocks before they hit the model. Most token burn in agentic coding comes from shoving raw shell output into context. Whether the 20–40% claim holds up in practice is unclear from the README, but the problem is real.

Key highlights

• Supports 40+ providers including free tiers (Kiro AI, OpenCode Free, Vertex’s $300 credit) and cheap ones (GLM at $0.6/1M tokens, MiniMax at $0.2/1M)
• Auto-fallback chain: subscription → cheap → free, with multi-account round-robin per provider
• Translates between OpenAI and Claude API formats so tools don’t care what’s upstream
• Dashboard for quota tracking and provider management at localhost:20128/dashboard
• Distributed via npm, Docker, and GHCR; source runs as a Next.js app

Caveats

• The npm package 9router is public, but the source repo package is private (9router-app), so running from source requires building the Next.js app yourself
• Several free tiers mentioned (iFlow, Qwen, Gemini CLI) are noted as “discontinued in 2026” — free provider availability shifts fast
• Token savings claims lack methodology or benchmarks in the README

Verdict

Worth a look if you’re running multiple AI coding tools and bleeding money on overlapping subscriptions. Skip it if you already have a single provider that meets your needs or if you don’t want another local service to babysit.