A CLI-based AI coding agent from Charm (Charmbracelet, Inc.), the company known for its polished terminal applications. It lets you delegate code generation, editing, and investigation to an AI without leaving the terminal. Its defining trait is that it is not tied to a single AI model: you connect your preferred LLM ─ OpenAI, Anthropic, Google, or a local model ─ and you can even swap models mid-conversation. Written in Go, it runs on macOS, Linux, Windows, and the BSDs. It descends from the Open Code lineage the same team worked on, and its source code is published on GitHub.
Key Features
- Bring your own LLM: Supports a wide range of providers including OpenAI, Anthropic, Google, Groq, OpenRouter, Hugging Face, and local models via Ollama. It follows a BYOK (Bring Your Own Key) model, so the tool itself costs nothing to use
- Switch models mid-session: Move to a different LLM while preserving conversation context. You can use a cheaper model for research and a more capable one for implementation
- LSP integration for accurate code understanding: Pulls type information and definition locations from the same Language Server Protocol your editor uses, so suggestions reflect the actual structure of the project
- Extensible via MCP servers: Supports the Model Context Protocol over stdio, HTTP, and SSE, letting you attach external services and in-house tools as extensions of the agent
- Per-project session management: Keeps multiple work sessions per project, each with its own persistent context, making it easy to pick up where you left off
- Permission controls and exclusions: Manage allowlists and denylists for commands the agent may run, and use
.crushignoreto keep specific files out of the AI’s view - Broad platform support: Runs on macOS, Linux, Windows, Android, FreeBSD, OpenBSD, and NetBSD, and installs through Homebrew, npm, Winget, Scoop, Nix, and other major package managers
Pricing
| Plan | Monthly Price | Highlights |
|---|---|---|
| Crush itself | $0 | Open source (FSL-1.1-MIT). No usage fee; you only pay the LLM provider you connect |
| Charm Hyper (Free plan) | $0 | Charm’s own model service. Includes 100 hypercredits that refresh monthly |
| Charm Hyper (Paid plans) | Check official site | Larger monthly credit allocations and priority access to models |
| Additional hypercredits | Pay as you go | Buy bundles beyond your monthly allocation; purchased credits do not expire |
Pricing is current as of August 2026. Please check the official site and Charm Hyper for the latest details.
Because Crush itself is free, your actual cost depends entirely on which models you use and how much. With your own API key you pay that provider’s rates; with a Charm-operated service such as Charm Hyper, you spend credits instead.
Pros & Cons
✅ Pros
- Freedom to choose models avoids vendor lock-in, and you can try new models as soon as they ship
- The tool is free and open source, so costs stay limited to API usage
- LSP integration means edits reflect project structure rather than plain text completion
- MCP support lets you build custom workflows connected to internal systems and external services
- Exceptionally broad platform coverage, including the BSDs and Android terminals
- A highly polished terminal UI, as you would expect from a team that has been building CLI tools for years
⚠️ Cons
- Getting started requires obtaining API keys and writing configuration, which is a hurdle for complete beginners
- Since pricing is metered API usage, heavy use can end up costing more than a flat monthly service
- It is a terminal-only tool with no GUI, which may not suit people accustomed to editor-based completion
- Output quality varies considerably by model, so you have to judge results yourself
- The FSL-1.1-MIT license means commercial use that assumes a fully permissive license (such as MIT) requires checking the terms
Comparison with Similar Services
| Criteria | Crush | Claude Code | Gemini CLI | Aider |
|---|---|---|---|---|
| Provider | Charmbracelet | Anthropic | Community (OSS) | |
| Available models | Choice of many providers | Claude family only | Mainly Gemini family | Choice of many providers |
| Billing | Free tool + metered API / Charm Hyper | Subscription or API | Free tier + API | Free tool + metered API |
| LSP integration | Yes | Proprietary code analysis | Proprietary code analysis | Repository map approach |
| MCP support | Yes | Yes | Yes | Limited |
| Implementation language | Go | TypeScript | TypeScript | Python |
Who Is It For
- Developers who want to compare and use multiple models without committing to one AI vendor
- Individual developers who want to keep costs down by paying only for what they use with their own API key
- People with terminal-centric workflows who want to delegate to an AI without leaving the shell
- Teams that want to connect internal tools and proprietary data sources to the agent via MCP
- Anyone working on Linux, BSD, or Android terminals, where tool support is often limited
Summary
Crush answers a specific need head-on: you want an AI coding agent, but you want to keep the choice of model in your own hands. The tool is free and open source, and your cost is whatever the LLM you picked charges ─ a structure that fits well in a period when model generations turn over quickly. The trade-off is some upfront preparation: API keys to obtain and configuration to write. Start by handing it a small task using an API key you already have, or Charm Hyper’s free allocation, and see whether it fits your workflow.