Minima

An agent harness that automates model selection. Its recommendation engine sends every task to the right model — up to 80% less token spend than running everything on Claude

minima

cost-aware model routing · recommend → run → judge → feedback → memory

▸ Do a deep dive into the repo and find blockers for publishing

routed to gemini-2.5-flash · memory · $0.0014 · conf 94%

◆ ls → cli/ connectors/ core/ runtime/

◆ read package.json → "version": "0.0.0", "private": true

The main blockers for publishing are:

1 every package.json has "private": true — blocks npm publish

2 versions are all "0.0.0" — need a semantic version bump

▸ plan and apply the fixes

routed to gemini-2.5-pro · prior · $0.0056 · conf 88%

◆ edit packages/cli/package.json

- "version": "0.0.0"

+ "version": "0.1.0"

✔ 4 packages updated · "private" removed

done · $0.0108 · saved 61% vs all-Claude · success

▸ write the release notes for v0.1.0

routed to claude-haiku-4.5 · memory · $0.0003 · conf 96%

trivial task → cheapest capable model

◆ read CHANGELOG.md → unreleased → 0.1.0

✔ wrote RELEASE.md

## v0.1.0 — packages now public · versions 0.0.0 → 0.1.0

done · $0.0003 · 1 turn · success

The same routing engine, from the terminal or from your code.

Minima CLI

Run automated tasks straight from your terminal. Install once and let Minima pick the model for every command.

Minima SDK

Wire cost-aware routing into your own app. Same engine, called from your code, so every request lands on the model that fits.