What Crush Does
Crush is Charm's terminal-native AI coding agent — a successor to OpenCode rebuilt with the same Bubble Tea TUI craftsmanship that made gum, glow, and lazygit beloved. It runs as a single binary on macOS, Linux, Windows, FreeBSD, and Android, talks to OpenAI, Anthropic, Google, Groq, AWS Bedrock, Azure, and any OpenAI/Anthropic-compatible endpoint, and treats your shell, working directory, and Git history as the AI's workspace. Bring-your-own-key from day one, no subscription, no cloud account.
TUI Polish and Daily Ergonomics
The first thing you notice is how Crush feels. Charm's design language — soft borders, distinct accent colors, smooth scrolling diff panes, and a status bar that always tells you which model you are talking to — makes it the most visually pleasant terminal agent on the market. Sessions persist between launches, mid-session model switching takes one keystroke without losing context, and the scrollback handles long agent traces without the rendering glitches that plague rougher TUIs.
Daily ergonomics are similarly considered. File operations show colored diffs before applying, MCP servers can be added via a simple config block, and the LSP-aware code understanding pulls type information from your existing language server rather than re-parsing the codebase from scratch. For developers already running Charm tools elsewhere in their workflow, it slots in naturally; for everyone else, the polish alone is a strong reason to try it.
Model Flexibility and Cross-Platform Reach
Crush's model-agnostic posture is a deliberate counterpoint to Cursor and Claude Code's vendor-locked experiences. You can run a session on Claude Sonnet for hard reasoning, switch to Groq's Llama for fast iteration, and end on a local Ollama endpoint for offline editing — all without losing the conversation. The OpenAI/Anthropic-compatible endpoint adapter means LiteLLM, OpenRouter, and self-hosted gateways all work.
Platform reach is the other unusual differentiator. Android support is rare in this category — most competitors only target macOS and Linux — and the FreeBSD build matters for the small but devoted community of developers running it as a primary workstation OS. Combined with the BYOK pricing model, Crush is one of the few coding agents you can realistically deploy in environments where commercial subscriptions or US-cloud dependencies are dealbreakers.
Limits and Trade-offs
Crush is younger and narrower than the established terminal agents. Aider has a deeper Git integration and better handling of large refactors across many files; Claude Code ships with Anthropic's most recent training and tighter agentic loop quality; OpenCode (the project Crush evolved from) still has a wider community-built tool ecosystem. Crush's planning loop is solid for small-to-medium tasks but starts to drift on multi-day refactors that span fifty-plus files.
Documentation also lags behind the code. Several MCP integration patterns and advanced configuration options are only discoverable through the source or Discord, and the BYOK billing means you pay attention to your own provider invoices rather than seeing per-feature cost breakdowns inside the tool. None of this is a dealbreaker, but new users coming from Cursor's onboarding will notice the rougher edges.