cogshift

local-first AI router · in the workshop

Your AI, in the right gear.

One address for all your AI tools. Cogshift automatically routes each request to the right model for the task — efficient when it can be, capable when it needs to be.

You keep working normally; Cogshift handles the switching. You pay the bigger price only when it matters, and your prompts stay on your machine.

Keep using your existing tools. No model picker, no switching by hand — Cogshift handles the model choice in the background.

In the workshop. Cogshift is being built in the open and isn’t ready to download yet. Check back soon.

your tools editor agent chat app one address cogshift routes every request to the right model two lanes fast lane routine work deep lane hard work your tools editor agent chat app cogshift routes every request to the right model fast lane routine work deep lane hard work
one endpoint in → the right model out

how it works

Three moving parts.

No new habits to learn. Point your tools at one address and the router does the deciding.

  1. 01  /  run it

    One small program on your machine

    Your tools — editors, agents, chat clients — all point at one address.

  2. 02  /  sort it

    A judge classifies the request before it runs

    Every request lands in one of two piles — routine, or hard — without you being asked.

  3. 03  /  shift

    The work goes to the model that fits

    Routine work takes the fast lane on cost-effective models. Hard work shifts up to the deep lane. You never think about model choice again.

the two lanes

Fast when it can be. Deep when it must be.

The decision is made up front, before the worker model starts — not after a weak answer comes back.

Fast lane

cost-effective models

Most of what you ask for is routine. That work runs on cost-effective models, and the bill stays small.

  • Short, well-specified turns
  • Routine edits and questions
  • The default, most of the time

Deep lane

capable models

When the judge classifies a request as hard, it shifts up to a capable model — and you pay the bigger price only then.

  • Open-ended, tangled problems
  • Work that needs real reasoning
  • The exception, not the rule

how it’s built

An appliance, not a middleman.

Local first

The router runs on your machine. Your prompts go to the models you chose — never through us.

One address

Editors, agents and chat clients all point at the same endpoint. The router decides what runs behind it.

Honest by design

Capabilities are detected, never guessed. What a route can’t do, it says it can’t do.

Budget, stretched

The strong model runs only when the work calls for it — and the counter proves it.

the ladder

Choose your gear.

Start with the router on your own machine and move up only as far as you want to — each step hands over one more thing to us.

01  /  start here

Download

free · bring your own keys

The router on your machine. Your keys, your models, your prompts — nothing leaves home, and nothing costs anything. This is the whole product for most people, and it is the one we are building first.

Download is in the workshop; hosted tiers follow.

where this is going

  1. 02

    Hosted

    your own router, run by us

    Your personal router at your-name.router.cogshift.io. Bring your own keys; reach it from every device.

  2. 03

    Hosted +

    routing models included

    We provide the judge — routing works out of the box, no configuration.

  3. 04

    Complete

    models included · pay as you go

    Workers included too. One bill, one address, every request in the right gear.

closed beta

Cogshift at work.

Cogshift for teams: the same router, wired into how your organisation already runs — and reporting on how it really uses AI.

Read about enterprise
  • Your identity providerPeople sign in the way they already do.
  • Per-user keys, with controlsIndividual keys, with controls and budgets attached.
  • Sensitive-data filteringBuilt in, with reporting on what it catches.
  • Business-intelligence reportsHow your organisation really uses AI.

questions

Straight answers.

What exists today, what doesn’t yet, and where your prompts go.

Where do my prompts go?

With the download, the router runs on your machine and talks directly to the models you configured. Your prompts stay on your machine and never pass through us.

What decides fast lane or deep lane?

A judge classifies the request before it runs — routine, or hard. Routine work takes the fast lane on cost-effective models; hard work shifts up to the deep lane. It happens per request, automatically; you never pick a model.

Which tools work with it?

Anything you can point at a different address — editors, coding agents, chat clients. They all use one endpoint, and the router decides what runs behind it.

Do I bring my own keys?

On Download and Hosted, yes — your keys, your models. Hosted + adds the routing models, so the judge is provided. Complete includes the worker models too, billed as you go.

Can I download it yet?

Not yet. Cogshift is in the workshop, being built in the open. The download comes first; the hosted tiers follow.

What about my team?

Enterprise is in closed beta: your identity provider, per-user keys with controls and budgets, sensitive-data filtering and reporting, and business-intelligence reports on how your organisation really uses AI.

get in touch

Say hello.

Cogshift is being built in the open. If you want to follow the build, ask about the enterprise beta, or tell us what you’d need from a router, leave a note.

There is no account to create and no list to join — this form only sends us a message. We read them, and we reply.