Crush: Glamourous AI coding agent for your favourite terminal

Original Article ↗ Hacker News Discussion ↗

Tool landscape & comparison difficulty

  • Commenters struggle to compare Crush with Claude Code, OpenCode, aider, Gemini CLI, Cursor, etc.
  • Several note that “which is best” depends heavily on model, codebase, and task; evaluation is a combinatorial explosion of tool × model × context × prompt.
  • Academic-style benchmarking is seen as expensive and skewed toward commercial models; some argue journals should de‑emphasize comparisons to opaque APIs.

Crush vs OpenCode and other agents

  • Crush is Charm’s rebranded fork of an earlier “OpenCode” effort after a high‑profile community dispute; this history is rehashed and remains contentious.
  • Direct user comparison to sst/opencode:
    • Pros for Crush: “sexy” TUI, nice diff view, good context display, LSP integration, clear Go codebase seen as a good blueprint for agents.
    • Cons: no Anthropic SSO, no GitHub Copilot auth, weaker planning/agent behavior, slower, higher token usage, junk binary artifacts, rough edges (history, editor, Ctrl‑C crashes). Many call it “beta” compared to OpenCode.
  • Others are bullish: they like Charm’s DX track record, Bubble Tea–based UI, FreeBSD/Go support, and early but active development.

Local models, endpoints, and “openness”

  • Strong interest in using local models (Ollama, LM Studio, llama.cpp, vLLM, sglang, etc.) to avoid cloud costs.
  • For Crush, local use is already possible via editing providers.json; first‑class Ollama/custom endpoint support is in progress or requested.
  • Experiences differ: some say “most agents work with OpenAI‑compatible endpoints,” others report real friction, especially with OpenCode GUIs and Ollama/tool‑calling.
  • Several note Crush is under the Functional Source License with a future MIT fallback; some expected fully open source and feel misled.

Terminal TUIs vs IDE workflows

  • Big split:
    • IDE fans (VS Code/JetBrains) see terminal agents as redundant, harder to integrate, and missing basic REPL affordances (scrollback semantics, selection, copy/paste).
    • Terminal/TUI fans value consistency across editors and SSH, lower resource usage, high information density, Unix‑style composability, and nostalgia for colorful TUIs.
  • Some prefer “plain CLI” agents like Aider that behave like a traditional REPL; others like richer, “glamorous” TUIs despite their quirks.

Agentic behavior, standards, and subscriptions

  • Users contrast Aider’s “single‑request” style with more autonomous agents like Claude Code (self‑planning, tests, iteration).
  • MCP support is considered important by some; Aider is criticized for lacking it.
  • There’s a push to standardize project instructions in AGENT.md instead of tool‑specific CLAUDE.md/CRUSH.md files.
  • Many want agents that can honor existing subscriptions (Claude Max, Copilot) instead of requiring separate per‑token API keys; OpenCode reportedly does this, Crush currently does not.