Crush vs Claude Code: Open Source vs Pro AI Coding Agent (2026)

The terminal-based AI coding agent space just got a lot more interesting.

Crush — a powerful open-source AI coding agent built in Go by the Charmbracelet team — has been taking Hacker News by storm with its promise of multi-model, terminal-first agentic coding. It's the active successor to the archived OpenCode project, built on the same ideas but as a fresh codebase.

Claude Code is Anthropic's official terminal agent, tightly integrated with Claude models and backed by the company behind some of the most capable AI models today.

Both run in your terminal. Both can read, write, and execute code. But they're fundamentally different tools. Here's the breakdown.

---

At a Glance

| Feature | Crush | Claude Code | |---------|-------|-------------| | License | MIT (Open Source) 🔓 | Proprietary (free tier + paid) | | Model support | Multi-model (OpenAI, Claude, Gemini, Groq, OpenRouter, local, and more) | Claude only (Sonnet 4 / Opus) | | Language | Go (via Charm TUI stack) | TypeScript (Node.js) | | Interface | Terminal TUI (Bubble Tea) | Terminal (CLI + input) | | GitHub Stars | ⭐24.7K | N/A (not open source) | | Cost | Free + your API keys | Free tier + usage billing or $20/mo subscription | | Setup time | ~2 minutes (brew/npm/winget) | ~3 minutes (npm + API key) | | Session persistence | SQLite-based workspace sessions | In-memory sessions (context window) | | LSP integration | ✅ Built-in | ❌ Not available | | MCP support | ✅ http, stdio, sse | ✅ (limited) | | Extensibility | ✅ MCP-based skills | ✅ Hooks system + skills | | Multi-client sessions | ✅ Multiple terminals, same workspace | ❌ Single session per terminal | | Platform | macOS, Linux, Windows, Android, BSD | macOS, Linux, Windows (WSL) | | Current maintainer | Charmbracelet | Anthropic |

---

1. The Origin Story: OpenCode → Crush

This is worth understanding because it explains the momentum behind the project.

March 2025: OpenCode launches on GitHub as a Go-based terminal AI agent. It gains traction quickly with its clean Bubble Tea TUI and multi-model support.

Early 2026: OpenCode reaches 12K+ stars. The community loves its flexibility — you can use Claude, GPT-4o, Gemini, or even local models with the same tool. The original project is later archived.

April/May 2026: The original authors, now working with Charmbracelet — the team behind Bubble Tea, Glamour, Lip Gloss, and the Go TUI ecosystem — launch Crush as a fresh codebase building on the same ideas.

May 2026: Crush hits 24.7K stars. The Hacker News discussion generates 1,274 points, making it one of the most upvoted AI tool launches of the year. Meanwhile, the original OpenCode repo remains online as an archived reference.

Claude Code launched in early 2025 as Anthropic's official terminal agent, and has been steadily growing. Its tight integration with Claude models makes it the default choice for anyone already using the Anthropic ecosystem.

---

2. Model Support: The Biggest Difference

This is the single most important factor in choosing between them.

Crush: Multi-Model

Crush supports a wide range of providers out of the box:

| Provider | Models | |----------|--------| | OpenAI | GPT-4o, GPT-4o-mini, o-series | | Anthropic | Claude Sonnet 4, Opus | | Google Gemini | Gemini 2.5 Pro, Flash | | Groq | Llama 3, Mixtral | | OpenRouter | 200+ models from all providers | | Charm Hyper | Dedicated inference endpoint | | Local endpoint | Ollama, LM Studio, vLLM | | Others | Azure, Hugging Face, Cerebras, and more |

You can even switch models mid-session while preserving context — a unique capability.

Claude Code: Claude-First

Claude Code is designed specifically for Claude. You can use it with an Anthropic API key, a subscription, or through certain third-party providers. The experience is optimized for Claude Sonnet 4 and Opus — and it shows. Claude Code's understanding of code context, multi-file reasoning, and command execution is deep because it's built for a single model ecosystem.

The tradeoff: If you want to experiment with GPT-4o or Gemini, you'll need a different tool. If you're all-in on Claude, Claude Code gives you the best possible experience with that model.

---

3. Cost Comparison

| Scenario | Crush | Claude Code | |----------|-------|-------------| | Heavy daily use | $20–80/mo (API costs vary by model) | $20/mo subscription (fixed) | | Light use | $5–15/mo | Pay-per-call or $20 flat | | Using GPT-4o only | Same API costs | Not possible | | Using local models | Free (no API cost) | Not possible | | Team/CI usage | API-key based, scalable | Subscription per seat |

The math: If you use multiple models and switch based on task complexity, Crush can be cheaper — use Gemini Flash for simple tasks ($0.10/million tokens) and Claude Opus only for complex refactors. Claude Code's subscription is simple but inflexible.

---

4. Terminal Experience

Crush: Polished TUI

Built on Charmbracelet's Bubble Tea framework, Crush delivers a rich terminal UI:

- Session-based workflow — maintain multiple work sessions and contexts per project - Multi-client — multiple terminals can join the same workspace, sharing session state, LSP, and MCPs - LSP integration — Crush uses Language Servers for additional code context, just like an IDE would - MCP extensibility — add capabilities via MCP (http, stdio, and sse) - Cross-platform — macOS, Linux, Windows, Android, FreeBSD, OpenBSD, NetBSD

Claude Code: Minimalist and Fast

Claude Code takes a different approach — minimal UI, maximum functionality:

- No TUI distractions, just a prompt and output - Commands are typed, not navigated - Focus is on the AI's reasoning, not the interface - Hooks system for automation (pre/post command scripts) - Git-based diff for reviewing changes

Which is better depends on your style. If you love a polished terminal experience with multiple sessions and IDE-like features, Crush wins. If you want the fastest path from prompt to code, Claude Code is lean and direct.

---

5. Real Workflow Comparison

Both tools can accomplish similar tasks, but they approach them differently.

Crush

Crush is session-based. You start a session in your project directory, and Crush loads project context (including LSP data). Multiple terminal windows can join the same workspace, sharing the session state in real time — useful for pair debugging or monitoring long-running tasks.

Key workflow features: - Workspace sessions — shared message history, permission queue, LSP, and MCP state across clients - Live session status — see IsBusy (agent turn in flight) and AttachedClients (how many are viewing) - CRUSH.md support — project-specific context files that Crush fills with build commands, patterns, and conventions

Claude Code

Claude Code works like a CLI tool: type a prompt, get a response. It uses a fork-based conversation model — you can branch off from any point in the conversation tree.

Key workflow features: - Hooks system — run custom scripts before/after Claude Code actions - Git workflow — changes are managed through git, with review prompts before applying - Session forking — revisit and branch from any previous point in the conversation

---

6. When to Choose Which

| Situation | Best Choice | |-----------|-------------| | You want to use multiple AI models | Crush | | You're all-in on Claude | Claude Code | | You want granular cost control | Crush | | You need LSP / IDE-like features in terminal | Crush | | You prefer minimal, fast CLI | Claude Code | | Open source matters to you | Crush | | You need hooks/automation | Claude Code | | Team deployment at scale | Claude Code (subscription model) | | Multi-terminal collaboration | Crush (shared workspaces) | | Hobbyist / budget-conscious | Crush (bring your own key) | | CI/CD integration | Either works well | | Local model experimentation | Crush (supports local endpoints) | | Enterprise compliance | Claude Code (official support) |

---

7. Getting Started

Crush

# Homebrew (macOS / Linux)
brew install charmbracelet/tap/crush
# NPM
npm install -g @charmland/crush
# Windows
winget install charmbracelet.crush
# Set your API key
export ANTHROPIC_API_KEY="sk-ant-..."
# or
export OPENAI_API_KEY="sk-..."
# or
export GEMINI_API_KEY="..."
# Start coding
crush

Configuration is stored in $HOME/.config/crush/crush.json:

{
  "lsp": {
    "go": { "command": "gopls", "enabled": true },
    "typescript": { "command": "typescript-language-server", "enabled": true }
  }
}

> Tip: Crush ships with a built-in crush-config skill — you can often ask Crush to configure itself.

Claude Code

# Install Claude Code
npm install -g @anthropic-ai/claude-code
# Or via VS Code extension
# Authenticate
claude login
# Start coding
claude

---

Verdict: Which One Should You Use?

Choose Crush if: - You value model flexibility over single-model depth - You want an open-source tool you can inspect and customize - You bounce between different models based on task and budget - You need LSP-aware code context in your terminal - You want multi-client workspace sessions

Choose Claude Code if: - Claude Sonnet 4 / Opus is your primary model (it's the best for coding) - You want a polished, supported product with Anthropic backing - You don't want to manage multiple API keys - You need hooks-based automation - You prefer a no-nonsense CLI with minimal UI overhead

The pragmatic answer: Try both. Install Crush for multi-model experimentation (throw in a Gemini key for cheap tasks), and keep Claude Code for heavy lifting. Many developers are running both side by side.

---

Related guides: - Claude Code vs Cursor: Which AI Coding Tool Should You Use? - Claude Code vs GitHub Copilot: Head-to-Head Comparison (2026) - How to Install Claude Code in 5 Minutes - Setting Up Claude Code with OpenClaw: Step-by-Step - Multi-Agent Coding Workflow Setup - AI Coding Agents Compared: Claude Code vs Copilot vs Codex vs Gemini