9Router Deep Dive: Free AI Routing, Token Saving, and Provider Fallback for Coding Agents

Editorial note

This article is based on the GitHub repo, README, package metadata, source tree, npm metadata, releases/tags, recent issues, recent PRs, and provider configuration files researched on June 3, 2026. The README's free/unlimited language is treated as project positioning, not independently verified quota advice.

5.1 It is a gateway, not a model provider

9Router does not magically create a new foundation model. It sits between coding clients and upstream providers. The router translates request shapes, refreshes tokens where supported, manages API keys, routes to accounts, and decides what to do when the preferred provider fails or hits a configured limit.

That distinction matters for risk. If an upstream provider changes a stream event, rejects a schema, removes a free tier, or limits account usage, 9Router has to adapt quickly. Recent issues show exactly that kind of provider churn.

Claude Code / Codex / Cursor / Cline
  -> http://localhost:20128/v1
  -> 9Router endpoint settings
  -> combo / alias / provider account
  -> upstream model API
  -> translated stream back to client

5.2 Combos are the routing primitive

The README frames fallback as a three-tier chain: subscription first, cheap providers second, free providers last. In practice, that becomes a combo configuration in the dashboard. A combo is where you encode model preferences, fallback order, and which accounts are allowed to serve a request.

This is useful for long coding days because manual provider switching is friction. It is also dangerous if you do not understand each provider's terms, quota, latency, and model quality. A fallback chain can keep work moving, but it can also quietly change model behavior mid-task.

5.3 RTK compresses agent tool output before translation

The most concrete technical feature is RTK Token Saver. Tool results from coding agents are often huge: diffs, search results, trees, logs, numbered reads, and shell output. 9Router detects those structures and applies specialized compression filters before the request is translated to Claude, OpenAI, Gemini, Vertex, Kiro, or another target shape.

The README says the filter is safe by design: if compression fails or makes output larger, the original text is kept. That is the right default because lossy or broken compression of a diff can make an agent edit the wrong code.

tool_result: git diff / grep / tree / logs
  -> RTK filter selection
  -> compressed or original payload
  -> format translation
  -> upstream provider request

5.4 Format translation is the hard part

The project supports OpenAI chat completions, OpenAI Responses, Claude-style messages, Gemini, Cursor, Kiro, Vertex, Antigravity, Ollama-style paths, embeddings, image generation, speech, search, and web fetch routes. That is a large compatibility surface.

Recent bugs show the cost: assistant prefill behavior, Vertex limitations, provider-specific token parameters, GitHub Gemini stream assembly, and media model tests can all break differently. 9Router's value comes from absorbing those differences, but that means the repo has to track upstream changes constantly.

5.5 Local storage and logs are part of the threat model

A router like this stores or touches provider credentials, OAuth sessions, request logs, usage records, model aliases, and possibly raw prompts. The current source includes local database adapters; the README describes local data under a data directory. Treat the machine running 9Router as credential-bearing infrastructure.

That does not make the project unusable. It means you should configure logging intentionally, use least-privilege API keys where possible, avoid routing sensitive client data through accounts with unclear terms, and keep the dashboard port private unless you have deliberately deployed it behind authentication.

Our Take

Use 9Router if you want a local, hands-on gateway for routing AI coding tools across multiple providers and you are willing to manage provider risk carefully. Skip it for regulated production code paths until you have reviewed credential storage, logging, provider terms, and fallback behavior under your own policies.

Q: Is 9Router an AI model provider?

No. It is a local router and dashboard that sends requests to upstream providers, accounts, free tiers, subscriptions, or compatible endpoints you configure.

Q: What endpoint should I configure in my coding tool?

Use `http://localhost:20128/v1` and the API key shown by the 9Router dashboard, then choose a model alias or combo defined inside 9Router.

Q: What is RTK Token Saver?

RTK compresses common coding-agent tool outputs such as diffs, grep results, file trees, logs, and numbered reads before the payload is translated and sent upstream.

Q: Why do versions differ between npm and GitHub Releases?

At research time, npm/tags were ahead of the latest GitHub release entry. For debugging, check npm metadata, tags, releases, and the default branch together.

Q: Is it safe to use OAuth/subscription accounts through a router?

That depends on provider terms and your risk tolerance. The project includes risk notices for some providers, so treat this as a policy decision, not a purely technical setup step.

Q: Where does local data live?

The README describes local data under the configured data directory, and current source includes SQLite adapters. Protect that directory because it can contain credentials, usage records, and logs.

Q: What breaks most often?

Recent issues point to upstream provider schema changes, streaming assembly, model IDs, token parameters, media endpoint tests, and packaged runtime dependencies.

Primary Sources

• 9Router GitHub repo
• README
• License
• Root package
• CLI package
• Next config
• Provider config
• npm package
• 9Router homepage

Issues and PRs

• Antigravity and Claude Code prefill issue
• GitHub Gemini empty response issue
• Xiaomi stream stall issue
• Runtime dependency pruning issue
• Agent workflow stability PR
• Media endpoint test PR
• HuggingFace media suggestions PR

Release Sources

• GitHub release v0.4.63
• GitHub tags