I stopped picking between Claude Code, Codex and Gemini, so I built a shell that picks for me

Caspian — Tue, 06 Oct 2026 09:11:50 +0000

I pay for more than one AI coding tool. Most of you probably do too.

And every day I hit the same friction:

A quick "where is X?" question burns my strongest model.
A big refactor lands on a fast model and comes back half-done.
I hit my Claude Code 5-hour limit while my Codex plan sits unused.
I switch tools and lose all the context of what I was doing.

So I built Caspian: one terminal copilot that sits on top of the agents you already have and picks the right one for each request.

How it works

Caspian runs Claude Code, OpenAI Codex, Gemini and Cursor in the background and shows their answers in one session.

For every request it triages the task and routes it:

The rules are simple:

Quick questions go to the cheapest model.
Hard changes go to the strongest one.
Subscriptions come before API keys, so you use what you've already paid for.
/use codex pins a tool if you want control; /use auto hands it back.

Pre-flight check before big changes

Before a large change runs, Caspian shows an estimate: files touched, lines changed, turns, tokens and time, next to what the fast model would cost. You choose: run it, run it on the fast model, or cancel. Dollar amounts only appear when you're billed per token.

It fixes vague prompts

Type add a forgot pasword link to it and a cheap model rewrites it as:

Add a "Forgot password?" link to the login page in src/pages/login.tsx.

You pick whether to send the refined version, the original, or edit it.

One memory across every tool

Decisions and changes are written to .caspian/context.md, so when work moves from Claude Code to Codex, the context moves with it.

Usage limits at a glance

The header shows how much of each tool's plan you've used, including the 5-hour and weekly windows and reset times. No more surprise rate limits mid-task.

Safe by default

Questions are read-only. Shell commands stay off until you press Shift+Tab. Big changes ask first.

Try it

You need at least one AI coding CLI installed and signed in, plus Node 20.18+.

npm install -g @caspianmd/caspian
cd my-project && caspian

It's free and runs on Windows, macOS and Linux. Requests go to each tool's own provider, exactly as if you ran it yourself.

👉 https://caspian.md/cli/

I'd love feedback, especially on the routing. Which tasks do you think it should send where? Drop a comment!

DEV Community: Caspian