Pick OpenCode if you want an open source (MIT) agent that works with almost any model: its site lists 75+ providers, local models included, and you pay only for tokens or a $10 a month Go plan. Pick Claude Code if you want Anthropic's own agent on Claude models with a flat Claude subscription, plus built-in cloud features such as web sessions and scheduled routines. Both run in the terminal and share the same ideas (skills, MCP, subagents, a read-only planning mode), so switching later is not hard.
OpenCode and Claude Code at a glance
As of October 2026, from opencode.ai, its GitHub repository and claude.com.
| OpenCode | Claude Code | |
|---|---|---|
| License | Open source, MIT | Not open source (Anthropic, all rights reserved) |
| Price of the agent | Free | Needs a paid Claude plan (Pro $20 a month and up) or an API key |
| Models | 75+ providers via Models.dev, local models, GitHub Copilot and ChatGPT logins, OpenCode Zen | Claude only: Opus 5.5 by default, Sonnet 5.5, Haiku 5.5, Fable 5.1 |
| Model subscriptions | OpenCode Go $10 or Go Plus $40 a month (open models and a few others) | Claude Pro, Max, Team, Enterprise |
| Where it runs | Terminal UI, desktop app (beta), IDE extension, web, GitHub and GitLab | Terminal, VS Code, JetBrains, desktop app, web, mobile app, Slack |
| Planning mode | plan agent (Tab to switch) | Plan mode |
| Skills and MCP | Agent Skills, MCP servers, custom commands | Skills, MCP servers, plugins |
| Hooks | JS/TS plugins on events (tool.execute.before, session.idle, permission.asked) | 33 hook events; shell, HTTP, MCP tool, prompt or agent handlers |
| Cloud and scheduling | None built in; run opencode run from cron or CI | Cloud sessions, routines on a schedule, API call or GitHub event |
Open source and model freedom
OpenCode's pitch is choice. It is MIT-licensed, its site reports over 208,000 GitHub stars and 950 contributors, and it connects to dozens of providers, including local models. It also offers its own model services: Zen, a pay-as-you-go list of models the team has tested for coding (billed per token from prepaid credit, with some free models for a limited time), and Go, a $10 a month subscription for popular open coding models, with Go Plus at $40 for higher limits.
Claude Code works the other way around: one model family, tuned for the agent, and sold with the Claude plan you may already have. You can also point it at Amazon Bedrock, Google Cloud or Microsoft Foundry, but it always runs Claude.
One detail decides it for many people. OpenCode's provider docs say Anthropic explicitly prohibits using a Claude Pro or Max subscription through third-party plugins, and OpenCode no longer bundles such plugins (since version 1.3.0). So if you want Claude models on a flat subscription, Claude Code is the supported route. In OpenCode, Claude means an API key or Zen, billed per token.
Daily use: build and plan
Both open in your project folder and take plain-language tasks.
- OpenCode ships two primary agents you switch with Tab:
build, with full access, andplan, which is read-only, denies edits and asks before running bash. Ageneralsubagent handles broad searches when you mention@general. - Claude Code has plan mode for the same job: it explores and proposes before it touches files. See our Claude Code plan mode guide. It also has subagents and agent teams for parallel work.
For scripted runs, opencode run "..." and claude -p "..." both execute a prompt and exit, so either fits in a cron job or CI step.
Extending each agent
Both read Agent Skills and connect MCP servers, so a skill folder with a SKILL.md and a remote MCP server can serve both.
The difference is hooks. Claude Code has declarative hooks in settings.json: you name an event such as PreToolUse or Stop and a command to run, and the command can block an action. OpenCode does this with plugins: JavaScript or TypeScript modules in ~/.config/opencode/plugins/ or .opencode/plugins/ that subscribe to events like tool.execute.before (inspect or block a tool call), session.idle (the agent finished) or permission.asked. Plugins are more flexible; hooks are quicker to write. Our Claude Code hooks guide shows the hook side.
Briefkin uses both mechanisms: its installer adds a hook to Claude Code and a plugin to OpenCode, so sessions from either agent land on one board, with a ping when one waits on a permission.
Cloud work and scheduling
Claude Code has more built in here. You can start sessions on claude.ai/code or the mobile app, pull them back with claude --teleport, and create routines that run in Anthropic's cloud on a schedule, from an HTTP call, or on a GitHub pull request or release. OpenCode runs where you run it: your machine, a server, or a GitHub or GitLab workflow. It also exposes a server and SDK if you want to drive it from your own code.
Which to choose
- You want open source you can audit and fork: OpenCode.
- You want to mix providers, or use local models: OpenCode.
- You want the lowest bill on open models: OpenCode with Go or Zen's free models.
- You want Claude models on a flat plan: Claude Code.
- You want cloud sessions, routines and mobile hand-off without building them: Claude Code.
- You are comparing more options: see Codex vs Claude Code and our list of Claude Code alternatives.
Questions
Can I use my Claude Pro or Max subscription in OpenCode?
Not in a way Anthropic allows. OpenCode's own provider docs say Anthropic explicitly prohibits using Claude Pro/Max through third-party plugins, and OpenCode stopped bundling such plugins in version 1.3.0. Use an Anthropic API key, OpenCode Zen, or Claude Code itself.
Is OpenCode free?
The agent is free and MIT-licensed. You pay for the model: your own provider key, OpenCode Zen's pay-as-you-go credits (some models are free for a limited time), or the OpenCode Go subscription at $10 a month.
Does OpenCode support skills and MCP like Claude Code?
Yes. OpenCode reads Agent Skills, connects MCP servers, and has custom commands and agents. Its equivalent of hooks is a JavaScript or TypeScript plugin that subscribes to events.
Does OpenCode have a plan mode?
Yes. Press Tab to switch from the default build agent to the plan agent, which is read-only: it denies file edits and asks before running shell commands.