Crush

• Fixing a failing build or test suite from the terminal while the agent reads compiler and language-server diagnostics
• Comparing models on the same task by switching providers mid-session without losing the conversation
• Running a coding agent on local models through Ollama, LM Studio, or llama.cpp for code that should not leave the machine
• Splitting a larger change across background subagents, then checking on or redirecting each one
• Running scripted agent tasks in CI or shell pipelines with crush run
• Coding over SSH on a remote server, a BSD machine, or an Android phone in Termux

Crush is a good pick if you want one terminal agent that is not tied to a model vendor and that understands your project the way your editor does. Language-server context, MCP, Agent Skills, and hooks are all built in, the provider list is long and updates itself through Catwalk, and it runs on more platforms than most rivals, including native Windows, Android, and the BSDs. Charm ships releases about weekly, and v0.100.0 added background subagents and finer command controls. Know three things before adopting it. The license is FSL-1.1-MIT: fine for internal and professional use, but source-available rather than open source until each release turns MIT after two years. The power features need discipline: project config files run as code, and Sysadmin and yolo modes remove guardrails. And it stays in the terminal, with no IDE extension or cloud agents, so teams that want those should look at Claude Code, Codex, or OpenCode's desktop and editor front ends.

About Crush

Crush is a terminal user interface (TUI) built in Go on Charm's own Bubble Tea stack. You start it in a project, pick a model with ctrl+l, and work in sessions: each project can keep several sessions with their own context, and you can switch models mid-session without losing the conversation. Models come from Catwalk, Charm's open, community-maintained provider and model registry, which Crush refreshes automatically (this can be turned off or pointed at a fork for air-gapped setups). Built-in providers include Anthropic, OpenAI, Google Gemini and Vertex AI, Amazon Bedrock (Anthropic models), Azure OpenAI, OpenRouter, Vercel AI Gateway, Groq, Cerebras, xAI, Moonshot, MiniMax, Z.ai, Hugging Face, and OpenCode Zen and Go; any OpenAI- or Anthropic-compatible API can be added as a custom provider, and Ollama, LM Studio, llama.cpp, LiteLLM, and oMLX models are discovered automatically. Instead of pasting a key, you can also sign in with GitHub Copilot, a ChatGPT subscription (added in v0.94.1), or xAI Grok OAuth (v0.98.0). Charm also runs Hyper, its own paid inference service on open-weight models (described in Charm's August 2026 release notes as a $20/month subscription, with prepaid options), which the README presents as the quickest way to start but which is optional. What sets Crush apart is how much of the developer environment it brings into the agent: language servers feed it diagnostics, references, and symbols and back rename and symbol-replace refactoring tools, MCP servers connect over stdio, HTTP, or SSE (with OAuth), Agent Skills load from Crush, Claude, Cursor, and shared .agents folders, and hooks can run on tool calls, turns, sessions, and permission events. Configuration moved in July 2026 to a crushrc file written in Bash and run by a built-in interpreter, so the same config works on Windows. Tools ask for permission by default, apart from provably read-only shell commands such as git status; you can allow, deny, or block specific tools and commands. Plan mode (v0.95.0) is for working out an approach before any code changes. v0.100.0 (November 5, 2026) added background subagents you can check on, message mid-run, stop, or move to another model, plus a Sysadmin mode that can run any command on your system, and prompts for SSH and GPG passwords inside Crush. Crush also runs headless with crush run, and a client/server mode, served locally over a Unix socket by default (or HTTP/TCP) with an OpenAPI-described API, lets several terminals or an app that embeds Crush share one workspace; this is a local API, not a hosted Charm service. The tradeoffs: it is terminal-only, with no IDE extension, desktop app, or cloud agents documented; the license is source-available rather than OSI open source (each release becomes MIT two years after it ships); pseudonymous usage metrics are on until you opt out; and output quality depends on the model you connect.