Crush vs OpenCode: pricing, features and which to choose

Crush is a Go terminal coding agent made by Charmbracelet, the company behind Bubble Tea and other terminal tools. The repository was created in May 2025 and has grown to more than 28,000 stars on GitHub. It runs on macOS, Linux, Windows through PowerShell or WSL, Android, FreeBSD, OpenBSD, and NetBSD. The application is free to install, but it does not include unlimited model inference. Charm lists Hyper as its official subscription provider, while users can connect hosted APIs or local models. That split matters for budgeting because Crush is the client, while the selected provider normally bills model usage. The project uses the FSL-1.1-MIT license, with each version converting to MIT on its second anniversary.

The normal workflow begins in a project directory, where Crush opens an interactive terminal interface and asks for a model or API key. Users can maintain multiple sessions and preserve context while switching models. Language Server Protocol integration supplies code intelligence, while Model Context Protocol support adds tools through HTTP, stdio, or SSE servers. The agent can inspect files, propose edits, and run commands, with configuration controlling permissions. Crush supports OpenAI-compatible and Anthropic-compatible custom providers, alongside Anthropic, OpenAI, Gemini, OpenRouter, Bedrock, Vertex AI, and Groq. Its security documentation says crushrc is trusted shell code and warns that shell substitutions in crush.json run during configuration loading. Configuration review therefore belongs in the normal deployment process.

Crush combines LSP context, MCP extensibility, sessions, local-model support, and broad operating-system coverage in one terminal client. Its cost and model quality depend on external providers, while its FSL license restricts some competing commercial redistribution before conversion. Configuration can execute shell code, so teams need permission discipline and repository review. Aider focuses on repository editing with an open-source license, Claude Code is Anthropic's first-party coding agent, and GitHub Copilot CLI is more closely tied to GitHub accounts and Copilot plans. Those products offer clearer single-vendor operating models. Pick Crush if endpoint flexibility, local inference, and terminal portability matter more than one bundled subscription or conventional open-source licensing.

OpenCode is an open source coding agent for terminal, desktop, and IDE workflows. The public site reports more than 215,000 GitHub stars, 950 contributors, 20,000 commits, and more than 16 million developers monthly, figures the company can change. The repository identifies the software as MIT licensed. Current documentation distinguishes OpenCode 1 and OpenCode 2, which share the same command name and cannot coexist under default installation paths. Version 2.1.2 is dated October 21, 2026. OpenCode is best understood as a client and agent layer rather than one hosted assistant. It can inspect repositories, explain code, edit files, run commands, and keep multiple sessions. The MIT license permits commercial modification and distribution subject to its conditions.

A typical workflow starts in a project directory with the opencode command. The documentation describes an /init command that creates an AGENTS.md file for project structure and coding conventions. OpenCode separates a build agent from a plan agent: the build agent can edit files and run commands, while the plan agent is read-only by default and asks before bash commands. Sessions can use LSP integration, provider selection, and a general subagent. Developers can connect OpenAI, Anthropic, Google, xAI, GitHub Copilot, ChatGPT Plus or Pro, local models, OpenCode Zen, or OpenCode Go. The terminal, desktop, and IDE clients support the same provider choice. This workflow is adaptable, although the account, model, quota, and data terms depend on the selected route.

OpenCode's main tradeoff is choice. Provider selection can reduce lock-in, but it also makes billing, privacy, and troubleshooting less predictable than a single-vendor product. The Go privacy table says most listed models do not use prompts for training, while Muse Spark models permit training and Gemini 4 Argon Contributor may use prompts and completions for future Google models. Compared with Claude Code, OpenCode offers local models and a broader provider route. Compared with Cursor, it is more terminal centered and MIT licensed. Compared with Aider, it adds desktop and IDE clients. Pick OpenCode for repository work when provider flexibility matters more than one unified account and bill.