What the comparison actually says
Jonker's main complaint about Claude Code on the desktop is that it lives as the Code tab inside the Claude app and only runs Claude models. He says even its third-party mode accepts only gateways that serve Claude models. So a trivial job, like renaming a folder of files, still draws on the same usage pool as his regular chats, even when a local model on his own GPU could do it. He also says Anthropic blocks Pro and Max subscriptions from working in third-party tools. That means the subscription only pays off inside Anthropic's own apps.
OpenCode's pitch is the opposite. According to Jonker:
- The composer has a model picker on every prompt. Most of his chores go to local models served through LM Studio, and a Claude model can be selected when needed.
- OpenCode's own free models appear in the same list.
- The desktop app launched in beta almost a year ago for Windows, macOS and Linux. It moved to browser-style session tabs in July, and OpenCode 2.0 arrived in October.
- Since 2.0, the terminal, web and desktop apps connect to one background service. Jonker says the desktop app runs the same agent as the terminal version.
OpenCode's own documentation describes the project as an open-source AI coding agent available as a terminal interface, desktop app or IDE extension.
Version context: what 2.2.0 added
The plugin claim holds up. OpenCode's GitHub release notes show v2.2.0 shipped on October 27 with experimental creation, installation and sharing of desktop plugins, plus a refreshed desktop app experience. The same notes say plugins are managed under Settings › Extensions › Plugins. They also list a restored, opt-in Todo tool, and a Windows desktop update installer that the notes say takes about a second for future updates.
Third-party release trackers show the app kept moving afterward. One tracker says v2.2.1 made a searchable model picker the default in desktop and added a way to watch and stop background shells from the activity pill. A changelog aggregator listed v2.2.5 as the latest release when checked. Jonker's account is therefore a snapshot from the 2.2.0 era, not a current feature list.
The plugin feature is also explicitly experimental. According to OpenCode's plugin documentation, a plugin that exports ./gui-experimental can add panels, commands or settings to Desktop. The app loads plugins from the server it runs, so the documented route is to configure them in the global opencode.jsonc. Treat it as an early extension point, not a mature marketplace. The release notes also flag breaking changes for plugin and SDK authors, so anyone building on it should expect churn.
Skills and CLAUDE.md: where the article needs a correction
Two migration claims deserve care.
Skills. OpenCode's skills documentation confirms it searches Claude-compatible locations, namely .claude/skills/<name>/SKILL.md in a project and ~/.claude/skills/<name>/SKILL.md globally. So Jonker's point that his existing skills appeared without setup is consistent with the docs. The documentation describes skills as loaded on demand through a native skill tool, where the agent sees the available skills and loads one when needed. Jonker says you call them by typing @. The documentation doesn't mention that, so treat it as his experience, not a documented behavior.
CLAUDE.md. Jonker says OpenCode 2.0 reads only AGENTS.md, so CLAUDE.md content must be copied across. OpenCode's current rules page says something different. It describes Claude Code compatibility as a fallback: a project CLAUDE.md is used only if no AGENTS.md exists, and ~/.claude/CLAUDE.md is used only if no global ~/.config/opencode/AGENTS.md exists. The page also gives environment variables to turn this off, such as OPENCODE_DISABLE_CLAUDE_CODE=1.
Per the same page, precedence is simple: the first matching file wins in each category, so if both AGENTS.md and CLAUDE.md exist, only AGENTS.md is used. My research notes say a V2 migration guide advises moving fallback guidance into AGENTS.md. I could not review that guide directly. The practical advice:
- Check whether your project has both
AGENTS.mdandCLAUDE.md. If so, only theAGENTS.mdcontent applies. - If you only have
CLAUDE.md, test whether your OpenCode version actually applies it. - For predictable behavior across versions, copy or consolidate the guidance into
AGENTS.md. - OpenCode's
/initcommand can generate anAGENTS.mdby scanning your repository.
The cost argument needs framing
The headline says "for free," and the software is. The models are a different matter.
- Local models. The cost shifts to your hardware and setup. Jonker has a GPU and uses LM Studio. Not every Windows PC will handle the same jobs as well.
- OpenCode's hosted free models. OpenCode's model documentation lists several as free, but for a limited time. The page also notes different data-handling terms per model. For example, some free models may use collected data to improve the model, and some NVIDIA-hosted trial endpoints say not to submit personal or confidential data. Check these before sending proprietary code.
- Claude through the API. Anthropic's help center says paid Claude plans and the Claude Console are separate products, and the API is billed separately. A Pro subscription won't pay for API calls made from OpenCode. Jonker's claim that API rates come to less than half the monthly subscription is his own calculation. It depends on the model, token volume and workload, so it isn't a general guarantee.
The fair summary is that you trade a flat subscription and a polished single-vendor experience for control over which model handles which task. That can save money, but it can also cost more if you run heavy Opus-class workloads through the API.
Where the desktop app still falls short
Jonker flags a few limits.
- MCP setup. He says connecting and signing in to MCP servers works in-app with
/mcps, but adding one still needs a terminal command or config file. OpenCode's MCP documentation matches this. It showsopencode mcp add ...for setup, and says that if a server needs authentication you open OpenCode, run/mcps, select the server and sign in. The documentation also covers manual configuration under anmcpblock inopencode.jsonc. - Built-in browser. He describes a browser tab in the side panel, opened with Ctrl+Alt+B, which the agent can use to read console errors on a page. The documentation I reviewed confirms a browser capability tied to the desktop app. I couldn't verify the shortcut, so treat it as version-specific. A release note for v2.2.0 also mentions keeping shortcuts working in the browser pane. A user-reported issue described the browser pane failing to appear on one Windows 11 installation of Desktop 2.2.2. That is a single report, not proof of a widespread bug.
- Polish. Jonker himself says Claude Code's app feels more polished, and he calls OpenCode's desktop app barely out of beta.
Practical takeaways for Windows users
- Install: OpenCode's docs list Chocolatey, Scoop, npm and Mise for Windows, and a release binary is available. The docs note that the install script and some methods install OpenCode 1 and can conflict with or replace an OpenCode 2 installation, so check which major version you are getting.
- Split the workload: Do what Jonker does. Route small, repetitive chores to a local or free model, and keep a frontier model for hard problems.
- Mind the data policy: Free hosted models differ in retention and training use. Read the model's terms before pointing one at private code.
- Don't assume subscription portability: A Claude Pro plan covers Anthropic's own apps. Using Claude in OpenCode means a separately billed API key.
- Pin your instructions: Put shared rules in
AGENTS.mdso they behave the same whichever tool, or version, reads them.
The broader trend is clear. AI coding agents are splitting into vendor-locked apps that polish one model's experience, and open front ends that treat models as interchangeable parts. Jonker's own verdict is to use both: OpenCode for the boring jobs, and Claude Code where he wants Opus on his subscription. For many Windows developers, that hybrid is probably the realistic answer.
References
- I ditched Claude Code's desktop app for its open-source rival that does the same jobs for free XDA · 2026-11-05T21:00:21+00:00
- Why is Claude API usage billed separately from my paid Claude plan? | Claude Help Center support.anthropic.com
- Agent Skills | OpenCode docs.opencode.ai