Crush

Pros: Crush is really appealing from a developer’s point of view because it brings agentic coding into the terminal without making the workflow feel bloated. It feels like a tool for people who already live in the command line and want an AI agent that can work with their existing code, tools, and habits.

I like that it is open source and model-flexible. Being able to wire your own LLM into the workflow is a big deal because different coding tasks need different tradeoffs around cost, speed, context, and reasoning quality.

The terminal UI is also a nice fit for real development work. Crush is not trying to replace the whole IDE; it gives you an agentic layer where you can ask for help, work through code, run commands, and stay close to the environment you already use.

The active GitHub repo is a good sign too. Frequent commits, lots of stars, and ongoing work around providers, MCP, hooks, and shell features make it feel like a project with real momentum instead of a one-off AI demo.

Cons: Crush is probably not the best fit for someone who wants a polished, hand-holdy coding assistant with everything abstracted away. It feels more geared toward developers who are comfortable in the terminal and want control over models, config, permissions, and workflows.

I would also be careful with any AI coding agent that can interact with your shell and project files. Recent research on coding-agent setup workflows shows that agents can be vulnerable to malicious or misleading project instructions, dependency names, registry redirects, and unsafe install steps, so developers still need review habits and guardrails.

For production repos, I would keep permissions tight, review commands, watch file diffs, and avoid letting any agent run destructive operations without explicit approval.

Overall: Five stars from me. Crush feels like a strong AI coding agent for developers who want something terminal-native, open source, customizable, and close to their real workflow.

It is not a magic autopilot, and I would not run it recklessly. But as a supervised coding partner for debugging, refactoring, repo exploration, command-line workflows, and agentic development, Crush looks like one of the more interesting tools in the AI coding agent space.