9Router

Pricing

Open source

Platforms

Localhost, Docker, VPS, Cloudflare Workers

Website

https://github.com/decolua/9router

Free access note

9Router can route coding tools toward free or low-cost provider paths such as Kiro AI, OpenCode Free, OpenRouter-compatible models, and custom providers.

Caveat

Free providers, quota terms, and model names can change quickly; verify the current README and provider dashboards before relying on them.

My take

9Router is interesting because it attacks a practical pain point in AI coding: quota burn and provider switching. Claude Code, Codex, Cursor, and Cline can all become expensive when logs, diffs, grep output, and long tool results are sent back to the model repeatedly.

My view: 9Router is best evaluated as infrastructure, not as a coding assistant. If it saves tokens, tracks quota, and keeps a session alive through fallback routing, it can make the existing agent stack cheaper and more resilient.

What it actually does

9Router exposes a local OpenAI-compatible endpoint, commonly http://localhost:20128/v1, then routes requests from coding tools to different model providers. The GitHub README describes support for Claude Code, Codex, Cursor, Cline, OpenCode, Antigravity, Copilot, Gemini, and other tools.

The most important feature is the RTK Token Saver. The README says it compresses tool outputs such as git diff, grep, find, ls, tree, and log dumps before the request reaches the LLM. That matters because these outputs often contain repeated filenames, unchanged lines, boilerplate logs, and long directory listings.

Why the 3-tier fallback matters

The strongest product idea is the fallback chain:

• Tier 1: subscription models for your preferred paid or quota-based access.
• Tier 2: cheap models when the main quota fails or becomes too expensive.
• Tier 3: free models when you need the coding session to keep moving.

For heavy AI coding, that is more useful than a generic model picker. The goal is not only to choose a model; it is to avoid having development stop when a provider returns an error, rate limit, or exhausted quota.

Where I would use it

• A Claude Code setup that frequently sends large terminal logs or diffs.
• A Codex or Cursor workflow that needs fallback providers behind one endpoint.
• A team experiment where cost needs to be measured before standardizing on paid models.
• A personal coding rig that mixes subscription, cheap, and free model paths.

Where I would be careful

9Router sits between your coding tool and model providers. That means it can see prompts, tool outputs, and potentially sensitive repository context. I would review logging, deployment location, provider credentials, and team policy before using it on private production code.

I would also avoid assuming every free provider is stable. Free access is useful for evaluation and fallback, but production coding workflows need predictable latency, terms, and auditability.

Quick start shape

The README shows the simplest path:

npm install -g 9router
9router

Then connect a provider in the local dashboard, point Claude Code or another tool at http://localhost:20128/v1, copy the dashboard API key, and choose a model or combo such as a Kiro, OpenCode, OpenRouter, or provider-backed route.

Coding Agent Tools verdict

9Router is worth a serious look if your AI coding workflow is already heavy enough that model cost, rate limits, and token waste are visible problems. It is less important for occasional prompting. The best adoption test is simple: put it behind Claude Code for one week, track token savings and fallback events, and decide whether the extra infrastructure is paying for itself.