Crush

Crush comes from Charm, and it shows: the interface is the nicest terminal you will use this year, and it wires language servers and MCP servers into whichever model you want. What you will notice day to day is that the agent sees what your editor sees, so it stops guessing at types and starts fixing the error you have. What you will not love: no sandbox, and a source-available license you should read before you fork.

Pick it for context quality on a typed codebase with your own keys. Pick OpenCode for a permissive license and subagents, and a larger crowd to file bugs with.

The flag is called --yolo. It disables the asking, and there is no Docker sandbox, so a headless run is an agent with your shell and nobody at the keyboard, running whatever the model decided a fix looked like. The name is honest, which is the only defense it has: nobody typed it by accident. Keep it interactive on anything that matters, and on anything else run it in a container you built.

What it does right is context: language servers feed the model diagnostics and symbols, so it reads the code the way the editor does rather than as text, and fewer of its guesses are wrong.

Crush makes one decision most peers do not: context is gathered through language servers, which provide typed symbols and diagnostics, rather than through lexical search alone, so the model receives the compiler's view of the file rather than a grep hit. Edits are applied to disk and verification is the shell. No benchmark is published, so the value of typed context is argued from architecture, not measured.

The claim is testable: the same task set with the language server disabled would isolate the effect. The observation: the autonomy setting is a flag named for exactly what it does, which is rarer than it should be.

This is Charm's move up the stack from terminal libraries to a product. Two tells. The license is source-available with a delayed open-source conversion, open enough for trust and closed enough to stop a hosted competitor, the standard playbook for a company that has watched clouds resell open code. And the Hyper provider at $20 a month sits next to free bring-your-own-key, so revenue comes from convenience, not lock-in.

Likely acquirer: whoever wants Charm's whole terminal stack and the audience it built with free libraries. Position: adopt for individuals, keep the license on file, and watch whether Hyper becomes the default.

The demo knows what the language server knows, which is the demo our senior engineers like. Procurement: no seat, tokens only, and Bedrock and Vertex are built in, so it runs on our existing cloud contract with no new vendor review and no new retention question. The license is a question for legal, since source-available is not a term the standard questionnaire has a box for. No sandbox and an autonomy flag mean a written laptop policy.

Onboarding is a brew install and a provider setting. Approved with conditions: keys via the cloud contract, the autonomy flag off by policy, and legal's read of the license on file.

Crush does two things I have wanted. It uses my language servers, so the model sees symbols instead of grep, and it auto-discovers local Ollama and llama.cpp, so the first run on my box needed no configuration. MCP servers go in a JSON mcp block, HTTP or stdio, with oauth as a field. The grudge is the license: FSL-1.1-MIT means a fork that competes with Charm is not allowed until the MIT conversion, which is a fork with a waiting period.

That is not open source, and I say so while running it daily. Grudging respect for the local-model discovery; a wary eye on the calendar.