Crush
The glamorous AI coding agent for your terminal — your tools, your code, your LLM of choice.
About
Crush is Charm's terminal-based AI coding agent — the company behind Gum, Bubble Tea, and Lip Gloss turned its famously beautiful TUI craft toward agentic coding. It runs entirely in your terminal on macOS, Linux, Windows, FreeBSD, and even Android, and talks to whatever model you prefer: Claude, GPT, Gemini, or any OpenAI/Anthropic-compatible endpoint.
The workflow is session-based: each project gets its own persistent session and context, and you can switch models mid-session without losing the thread. Crush leans on language servers (LSP) for code understanding rather than naive text dumps, and it extends through MCP servers over stdio, HTTP, and SSE. Tool execution is permission-gated by default, so it asks before running shell commands.
With ~28,600 GitHub stars it is one of the fastest-growing coding agents of 2025–2026, and being MIT-licensed with no account required makes it the pragmatic pick if you want Claude-Code-style agentic coding without vendor lock-in. The honest caveat: it is young and moves fast — expect occasional rough edges and breaking config changes between releases.
What it can do
- Multi-model: Claude, OpenAI, Gemini, or any OpenAI/Anthropic-compatible API, switchable mid-session
- Session-based organization with persistent per-project context
- LSP-enhanced code understanding, like a human developer's editor
- MCP extensibility over stdio, HTTP, and SSE transports
- Permission-gated tool execution (configurable)
- First-class terminal UI from the Charm ecosystem (Bubble Tea); MIT licensed
Setup
- Install via a package manager:
brew install charmbracelet/tap/crush,npm install -g @charmland/crush,winget install charmbracelet.crush, or your platform equivalent. - Run
crushin your project directory; on first launch it prompts for provider API keys and a default model. - Approve or deny tool executions as it works — tune permissions in
~/.config/crush/. - Add MCP servers (stdio/HTTP/SSE) to the config to extend its capabilities.
Prerequisites
A terminal you like; an API key for at least one LLM provider (or a compatible local endpoint). No account or signup needed.
Use cases
Works with
Trust score — 9/10
Scored 9/10 (verified tier, metrics-v1). Evidence: Adoption 8: starScore from 28,663 stars on charmbracelet/crush — one of the fastest-growing coding agents; Maintenance 8: pushed 2026-10-15 (recency 10); frequent releases assumed 8; issues ratio 855/28663 -> 4; Manual 9: official site + repo verified, install paths documented; deduction: young project, fast-moving config. License: MIT.
★ 28,663 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 9.0/10
- · Adoption 8: starScore from 28,663 stars on charmbracelet/crush — one of the fastest-growing coding agents
- · Maintenance 8: pushed 2026-10-15 (recency 10); frequent releases assumed 8; issues ratio 855/28663 -> 4
- · Manual 9: official site + repo verified, install paths documented; deduction: young project, fast-moving config
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.