AI Expertise

Description

An AI coding agent for the terminal from Charm. It reads your project, edits code and runs commands, and asks for permission first; only commands that change nothing run without a prompt.

You choose the model: keys from many providers, addresses compatible with OpenAI and Anthropic, and local models, for example through Ollama. You can switch models mid-session and keep the context. Each project can hold several work sessions. To understand code, Crush connects to language servers (LSP), and you add new abilities through MCP and skills. The company has its own model provider, Hyper, with subscription and prepaid options, but it is optional.

It runs in the terminal on macOS, Linux, Windows, Android, FreeBSD, OpenBSD and NetBSD.

The source code is available, but this is not open source. Under the FSL-1.1-MIT license you may not release a commercial product or service based on it that competes with Crush. Internal use and non-commercial education and research are allowed. Each version moves to the MIT license two years after its release.

Pros and cons

Pros

  • Works with models from many providers and with local models
  • You can switch models mid-session and keep the context
  • Asks for permission before acting
  • Runs on many systems, including Windows and Android

Cons

  • Not open source: the FSL-1.1-MIT license bars competing commercial products based on the code
  • Terminal only
  • Models are paid separately unless you run them locally
  • Collects pseudonymous usage metrics by default; an environment variable turns this off, and the developer says prompts and responses are not collected
  • Config files run as code, so starting Crush in an unknown folder is risky

Key facts

Free access
Yes The program is free; models are paid to the provider you choose, and local models work too
Developer
Charm (Charmbracelet, Inc.)
Checked on
Nov 4, 2026

Editorial score

How strongly the AI Expertise editors recommend this tool for each of its categories right now, on a 0–10 scale. The score combines output quality, features, beginner friendliness and freshness.

Coding

8.2 / 10
Output quality 9.0
Low High
Features 8.0
Basic Rich
Beginner friendliness 5.5
Hard Easy
Freshness 10.0
Outdated Fresh