OpenCode vs Claude Code

OpenCode vs Claude Code is a search people type when they already like a terminal agent and want to know whether to stay on Anthropic or switch to an open harness. The useful split is not “which one writes more lines.” It is license, model choice, surfaces, and who is allowed to call the change finished.

OpenCode’s own intro calls it an open source AI coding agent, available as a terminal interface, desktop app, or IDE extension. You cd into a project, run opencode, optionally /init to write AGENTS.md, then Plan or Build. That is a real, documented product. This post does not steal its stars, and it does not pretend Critique ships it.

Claude Code’s official overview describes an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools. Anthropic ships it in the terminal (claude), IDE extensions, a desktop app, and the browser, with Slack, GitHub Actions, GitLab CI, and other surfaces documented. Most of those surfaces want a Claude subscription or an Anthropic Console account. The Terminal CLI and VS Code also document third-party provider paths.

OpenCode vs Claude Code at a glance

What OpenCode is, from its own docs

OpenCode is the open source AI coding agent at opencode.ai. The docs tell you to install, /connect a provider, cd into a project, run opencode, then /init so it can analyze the repo and write AGENTS.md. Plan mode (Tab) disables edits so the model can propose an implementation. Build mode makes the changes. /undo and /redo revert or restore the last agent edits. /share creates a link to the current conversation.

The agents page documents two built-in primary agents (Build and Plan) and three built-in subagents (General, Explore, Scout). You can add your own in opencode.json or markdown under ~/.config/opencode/agents/ and .opencode/agents/. A review-style subagent with edit: deny is a documented pattern. That is useful. It is still the same harness configuring a specialist. It is not an independent controller that fires after every claimed unit of work whether the author wants it or not.

Providers docs say OpenCode uses the AI SDK and Models.dev for 75+ LLM providers and local models. /connect stores keys in ~/.local/share/opencode/auth.json. Zen is optional: a handpicked set of models OpenCode has tested for coding agents, billed through opencode.ai/auth if you choose that path. You can also bring Anthropic, OpenAI, Google, Copilot login, ChatGPT Plus/Pro login, or a custom OpenAI-compatible endpoint. That model-agnostic story is the product’s honest strength, including if you want Claude models without living inside Claude Code.

Permissions default to allow for most tools. The docs are explicit about this. doom_loop (same tool call three times with identical input) and external_directory default to ask. .env files are denied on read unless you override. You can tighten everything to ask, deny git push, or run --auto so non-denied prompts auto-approve. Configure that if you care. Do not assume a fresh install behaves like a locked-down review gate.

What Claude Code is, from Anthropic’s docs

Claude Code is Anthropic’s agentic coding tool. The overview is blunt: it reads the codebase, edits files, runs commands, and plugs into development tools. Install with the official script, cd into a project, run claude. First use logs you in unless ANTHROPIC_API_KEY is already set. One-shot tasks use claude "task" or claude -p "query". claude -c continues the most recent conversation in the current directory.

The how-it-works page frames Claude Code as the harness around Claude: models reason, tools act, the loop feeds tool results back into the next step. That is the same category OpenCode occupies. The difference is vendor and surface area. Claude Code is built to follow you from terminal to VS Code, JetBrains, desktop, web, Slack, and CI. Sessions can move: Remote Control, /desktop handoff, claude --teleport from web or mobile into the terminal where those features are documented.

CLAUDE.md is the project instruction file. Anthropic also documents auto memory, skills for repeatable workflows, hooks that run shell around actions, sub-agents, and an Agent SDK. Git workflows are first-class: stage, commit, branch, open a pull request. That depth is why teams that already pay for Claude usually stay. Switching to OpenCode because a blog post used the word “open” is a bad trade if the surfaces you use are Claude Code’s.

Claude Code is not model-agnostic in the OpenCode sense. The default path is Claude. Cloud deployments go through Anthropic or through Bedrock, Vertex-class, and Foundry as the enterprise docs describe. Third-party providers exist on some surfaces. They are not the product’s center of gravity. If your constraint is “I must swap models every week without leaving the same MIT TUI,” OpenCode matches that constraint more honestly.

How to choose OpenCode or Claude Code

Stay on Claude Code if you already have a Claude subscription, you like CLAUDE.md plus the IDE and desktop apps, and you want Anthropic to keep shipping the harness. The cost is vendor concentration. The author can still tell you the tests were green.

Move to OpenCode if you need MIT source, a forkable TUI, Plan/Build as documented modes, and /connect across 75+ providers including local models. You can still point OpenCode at Anthropic. You are not required to. Star count is popularity, not a SWE-bench score. You still owe the repo a review step the author did not grade.

If the comparison you actually wanted was another open CLI against Claude Code, read Aider vs Claude Code. Aider is a different open-source terminal author (git-centric, pair-programming style) and should not be collapsed into OpenCode. The comparisons hub lists the other head-to-heads.

After you pick an author, add a sidecar if you need independent review

Both OpenCode and Claude Code are authors. Optional review subagents, Plan mode, skills, and CI bots are still configuration on the same product family. If your actual failure mode is “the writer graded its own homework,” keep the author you like and add an independent finish pass.

That sidecar is the Critique CLI: package @critiquedotsh/cli, binary critique. Install it next to OpenCode or Claude Code. Do not install CritiqueCode as a sidecar. CritiqueCode is @critiquedotsh/harness, binary critique-code. It is an optional author for teams that want Critique to implement in an interactive session, then force local review before promotion. Most people who already like OpenCode or Claude Code should keep that author and use critique. Product split: CritiqueCode docs. Intro: Introducing CritiqueCode.

FAQ