The right model on every job. Automatically.

AiAx Router is a free, open-source dispatcher for the AI subscriptions you already pay for. It splits your request into jobs, hands each one to the strongest CLI agent you have for that kind of work at the right reasoning effort, and has five review agents grade the result before you see it. You never pick a model. No API keys, no extra bill.

Download for Mac Download for Windows Open the repo

macOS 13+ Apple silicon · Windows 10+ · Intel Mac and Linux on the releases page · free under MIT

DISPATCH LOG · ONE TASK, REPLAYED

task"Fix the checkout rounding bug, then write the release note"

intentclaude opusxhighdefines what done means · cuts 3 jobs

job 1coding · hardcodex gpt-5.6-solhighpass

job 2tests · mediumclaude sonnetmediumpass

job 3writing · easyclaude haikulowpass

assemblethree pieces, one answer

reviewcorrect 10complete 9.5quality 10robust 9.5simple 10

verdict9.8 / 10 · shippedeasy work went cheap, opus stayed fresh

The point of it

Why not just open Claude Code or ChatGPT?

Because one app is one model at one effort level, and you are its review process. For everyday tasks and for coding, that wastes both quality and quota.

One app on its own Through the router
One model, at whatever effort you happened to pick. Each job gets the strongest model you have for that kind of work, at the effort the difficulty calls for.
Frontier quota burns on renames and one-liners. Easy parts go to cheap models. The expensive capacity is still there when the work turns hard.
You read the output and hope it is right. Five review agents grade every result for correctness, completeness, quality, robustness and simplicity. Under a 9 goes back before it reaches you.
You paste the context and instructions yourself, every time. Built-in skills ride along automatically, matched to the kind of work each job is.
Every subscription is its own island. Claude, ChatGPT, Grok and Kimi pool into one crew. Only Claude? It still spreads work across Opus, Sonnet and Haiku, so one plan goes further.

Decided by measurement, not habit.

Routing runs on public benchmarks that refresh weekly, and just as much on what a task actually costs each model in tokens. A model that scores slightly higher but needs twice the tokens loses the easy jobs and keeps the hard ones. That is the whole trick: quality per token, per kind of work, decided from real measurements instead of a favourite.

What happens when you press Enter

You write one line. The router runs the rest.

  1. 01Understand. The best model available works out what a great result would look like before anyone starts.
  2. 02Split. Big requests break into smaller jobs, each tagged with the kind of work it is and how hard it is.
  3. 03Dispatch. Every job goes to whichever agent measures best for that work, at the right effort, with the right skills. If one drops it, the next takes over.
  4. 04Assemble. The finished pieces come back and get merged into one answer.
  5. 05Inspect. Five review agents grade it. Under a 9, it goes back for another pass, and whatever still falls short is said out loud.
One real task, recorded in the app: typed in chat, split into jobs, dispatched with the right skill, then opened from the board.

No API keys anywhere

Your subscriptions do the work. The router keeps the receipt.

Every task totals what the same tokens would have cost at API list price, so you can see what your flat-rate plans just absorbed. The router talks to the official CLI agents you have already installed and signed in to. Nothing to top up, nothing metered.

Put your subscriptions to work.

Download for Mac Download for Windows Open the repo