Crush: review, pricing, security and alternatives

What Crush does

Crush belongs to the coding agents family: they read a repository, edit files, run the tests and fix things until everything passes, in a terminal or an editor. On our grid it ranks #19 of the 74 projects of this family listed in the directory.

You use it from: terminal (CLI). Usable models: Claude (Anthropic), GPT (OpenAI), Gemini (Google), OpenRouter, local models (Ollama, llama.cpp, vLLM…), and any OpenAI-compatible model. It can run on your own machine or server.

Who is Crush for?

It is aimed at developers and technical teams who want to hand off part of the work on their codebase while keeping the final review. Typical tasks people give it:

• fix a bug from an issue or an error message
• write or extend tests
• migrate code to a new library version
• explain an unfamiliar codebase

It is a good candidate if you are comfortable in a terminal, your data must stay on your machines, and you want the option of a local model, with no API fees.

Access and human control

According to its documentation, Crush can:

• run commands in a terminal
• read and edit files
• call online services and APIs

By default it asks for your confirmation before sensitive actions (command, send, delete). Keep that setting, at least at first: full-auto mode saves time but removes the last safeguard.

Crush pricing

The code is public but the repository declares no licence: reuse is not formally allowed. The real cost comes from model calls: each loop step sends the whole context back to the model, and a task of several dozen steps can cost far more than a single question. The LLM price calculator estimates it for the chosen model.

Getting started with Crush

1. Install it: the project ships as a single Go binary, downloadable from the repository releases. The exact command is in the repository README.
2. Pick the model: an API key from a provider, or a local model if your machine has enough graphics memory (the VRAM calculator tells you).
3. Give it a first playground with nothing at stake: a test folder, a secondary account or a throwaway project. Only open your real accounts once you have seen how it behaves.
4. Give it a short, checkable task, review the result and its action log, then raise the difficulty step by step.

Security: risk level

A coding agent runs commands on your machine: work on a Git branch, review the changes and isolate it in a container for long tasks.

Points found for Crush:

• Supervised sensitive access: it can act on the system, but asks for confirmation by default.
• Outside content: it reads web pages or API responses, which may contain booby-trapped instructions (prompt injection).
• No licence: legally shaky for company use.

Three habits: least privilege (minimal rights, dedicated keys), human approval for irreversible actions, a sandbox for code execution. The AI agent security check reviews your setup in a few minutes.

Project activity and releases

Created on May 21, 2025, the charmbracelet/crush repository has 28,794 stars, 2,322 forks and 309 open issues, with 4,350 commits in total. It received changes this month.

Over the last 52 weeks: 2,240 commits, an average of 43.1 per week.

196 releases in total; 29 in the last twelve months. Latest release: v0.98.2, on October 23, 2026.

Crush vs other coding agents

GitHub stars of the best-rated open-source projects in the coding agents family. Stars measure attention, not quality: a heavily promoted project collects them fast, a specialised tool will have few.

Crush and other coding agents at a glance

The 7 best-ranked agents in the same family, 7 of them open source. The score out of 100 only applies to open-source projects; for proprietary services we show the published entry price. A dash marks information the documentation does not state.

How we rated it

Score: 75/100, computed only from public, verifiable data, with no subjective testing or partnership:

• Adoption 30/40: 28,794 stars on a log scale (40 points from about 200,000).
• Activity 30/30: last commit less than a month ago, 29 releases in six months, 2,240 commits in a year.
• Framework 15/30: licence, default human approval, self-hosting, model choice and documentation site.

The score does not say whether the agent will succeed at your tasks: it measures how solid the project is. The full grid is on the method page.