Crush is a terminal-native AI coding agent built in Go by Charmbracelet, the team behind widely adopted TUI libraries such as Bubble Tea and the Glow markdown renderer. Released in early 2025, it brings the same polish Charmbracelet is known for to agentic coding workflows, offering a full-featured interactive TUI, LSP-backed code understanding, and MCP extensibility in a single native binary. With over 25,000 GitHub stars and support across Windows, macOS, Linux, Android, and BSD variants, it has rapidly become one of the most platform-inclusive options in the agentic CLI space.
Key capabilities
LSP integration for deep code understanding — Unlike most CLI coding agents that rely solely on file reads, Crush integrates with Language Server Protocol servers to give the model accurate symbol resolution, type information, and project-wide reference graphs, reducing hallucinations on large codebases.
MCP server support — Crush supports the Model Context Protocol as both a client and extensibility surface, allowing teams to wire in custom tools, internal knowledge bases, or third-party MCP servers without forking the tool.
Session management and —yolo mode — Crush persists conversation sessions so context survives across terminal restarts, and exposes a —yolo flag for fully autonomous execution where the agent applies file edits and runs commands without per-step approval, suitable for sandboxed or automated environments.
Broad provider and platform support — The tool ships with drivers for Claude, GPT-4o, Gemini, DeepSeek, and local models via Ollama, and its native Go binary runs without a runtime dependency on Windows, macOS, Linux, Android, and BSD variants.
Autonomy level
Crush operates at autonomy level 3 in its default interactive mode, confirming edits and shell commands with the user before executing them. When launched with —yolo, it advances to level 4, planning and executing multi-step tasks without per-action prompts. Neither mode reaches level 5 because Crush does not autonomously deploy to production systems or manage secrets. The —yolo flag is intended for use inside containers or CI pipelines where the blast radius of unreviewed changes is bounded.
Strengths
- Polished TUI from a team with a proven record of high-quality terminal tooling
- LSP integration gives the model richer code context than file-read-only agents
- Broadest platform coverage in its class, including Android and BSD variants
- Single native Go binary with no Node, Python, or runtime dependencies to manage
- MCP extensibility for custom tools and data sources
- Multi-provider support including local Ollama models for air-gapped environments
- FSL-1.1-MIT license converts to MIT two years after each release, providing a clear open-source path
Limitations
- FSL-1.1-MIT is source-available but not OSI-approved open source; commercial redistribution is restricted until the conversion window passes
- —yolo mode requires careful scoping; running it outside a sandbox against a production repository carries real risk
- Newer tool with a shorter track record than Aider or Claude Code in enterprise settings
- LSP integration requires a compatible language server to be installed and configured per project