# Open Claude Cowork: a desktop shell around the Claude Agent SDK > Open Claude Cowork (package name agent-cowork) wraps the Claude Agent SDK in an Electron desktop app for macOS, Linux and Windows. It is an early-stage project, and the README leaves more questions open than it answers. **DevAgentForge/Open-Claude-Cowork** — OpenSource Claude Cowork. A desktop AI assistant that helps you with programming, file management, and any task you can describe. - Repository: https://github.com/DevAgentForge/Open-Claude-Cowork - Stars: 3,396 · Forks: 459 - Language: TypeScript - License: not declared - Published: 2026-09-23 · Updated: 2026-09-23 · Language: en - Canonical page: https://hysenlabs.com/projects/devagentforge-open-claude-cowork ## What Open Claude Cowork is for, and who it is not for Claude Code runs in a terminal. The README states the problem plainly: no visual feedback for complex tasks, hard to track multiple sessions, and tool outputs that are inconvenient to inspect. Open Claude Cowork, whose package name is agent-cowork, is the project's answer to that. It is an Electron desktop application that drives the Claude Agent SDK and presents the work in a window. The intended audience is people who already use Claude Code and want a graphical layer over it, not people who want to avoid Claude tooling altogether. The README says the app reuses your existing ~/.claude/settings.json and that no development environment or Claude Code installation is required to run it. Those two statements sit awkwardly together, since the build prerequisites list Claude Code as installed and authenticated. Read the second claim as applying to the downloaded release, and the first as applying to both paths. If you want a coding agent embedded in your editor, this is the wrong shape of tool. It is a separate desktop process, not a plugin. ## The architecture visible in the repository The stack is Electron plus React plus Vite. package.json sets main to dist-electron/main.js, and the dev script runs two processes in parallel: a Vite server for the renderer and a compiled Electron main process. TypeScript is compiled separately for the Electron side through src/electron/tsconfig.json, which the transpile:electron script invokes. State in the renderer is handled by zustand, with Radix UI primitives for dialogs and menus and Tailwind for styling. Markdown rendering goes through react-markdown with remark-gfm, rehype-raw and rehype-highlight, which tells you the app expects to display model output containing tables, raw HTML and fenced code. That is the mechanism behind the visual feedback claim: tool output is rendered as formatted content rather than raw terminal text. Persistence uses better-sqlite3. The rebuild script targets that native module specifically, which is the usual sign that it needs recompiling against the Electron ABI. The dependency on @anthropic-ai/claude-agent-sdk is what actually executes agent work; the app is a front end around it, not a reimplementation. os-utils appears in the dependency list, suggesting some system resource reporting, though the README does not describe what it displays. ## Installing Open Claude Cowork and running a first task The README gives two paths. The first is to download a release from the project's Releases page. The second is to build from source. The build prerequisites are Bun or Node.js 22 or newer, plus Claude Code installed and authenticated. Clone and install dependencies: ```bash git clone https://github.com/DevAgentForge/agent-cowork.git cd agent-cowork bun install ``` Run the app in development mode with hot reload. The dev script kills any existing vite and electron processes first, waits a second, then starts the renderer and the Electron main process together: ```bash bun run dev ``` When it starts you should see a desktop window. The README does not describe the first-run screen, so what appears on launch is not documented. To produce a distributable binary instead, pick the target that matches your machine. The macOS Apple Silicon build is: ```bash bun run dist:mac-arm64 ``` The other targets are dist:mac-x64, dist:win and dist:linux. Note that dist:win exists as a script even though the README's platform badge lists only macOS and Linux, which is a discrepancy worth knowing about before you rely on the Windows build. If you hit a native module error from better-sqlite3, the repository provides a repair script rather than leaving you to guess: ```bash bun run rebuild ``` ## Where the documentation stops short The roadmap section contains a single word: todo. There is no list of planned features, no issue of what is unfinished, and no indication of which parts of the app are considered complete. For a project at version 0.1.0, that is a real gap. You cannot tell from the README whether session management, file operations, or the SQLite persistence layer are finished or partial. The README also never documents what happens when a task fails, whether sessions can be resumed after the app closes, or how the SQLite database is scoped per project. The example in the README is a video of organizing a local folder, which shows the shape of a task but not the interface around it. Contributing guidance is equally thin: fork, branch, commit, open a pull request, with a request to make only minimal changes. There is no test command in package.json at all, only lint and build. If you plan to modify the code, you are working without a test suite to check yourself against. One more inconsistency: the README badge says version 0.0.2 while package.json says 0.1.0, and the releases list includes 0.1.0 from 2026-01-19. The badge is simply stale. ## How it compares with Claude Code on its own The honest comparison is not with another desktop AI assistant, it is with the terminal tool this project wraps. Claude Code gives you a text interface, scriptable, composable with shell pipelines, and usable over SSH. Open Claude Cowork gives you a window with formatted output, which is genuinely better when you are reading long agent transcripts or tracking several sessions at once. The trade is real. A terminal tool can be piped, logged and run headless on a remote machine. An Electron app cannot. It also carries the usual Electron cost in memory and disk, and it depends on a native SQLite module that has to be rebuilt against the Electron version you are running, which is exactly why the rebuild script exists. If your work happens on a remote server, or if you script agent runs into CI, this project adds nothing and takes away composability. The README's framing of the project as a desktop alternative to Claude Cowork is accurate in shape but not in maturity. Claude Cowork is a shipped product; this is a 0.1.0 repository whose last push to main was on 2026-03-21. ## Licence and the cost of keeping it current The README's final section states MIT. That is a permissive licence, so you can use, modify and redistribute the code provided you keep the copyright notice and the licence text with it. One caveat belongs here: the top-level repository listing does not show a LICENSE file among the entries. The README says MIT, but if you need certainty for a commercial deployment, confirm the licence file exists in the tree you actually clone rather than relying on the README line. This is a factual observation, not legal advice. The upgrade cost is the interesting part. The project pins @anthropic-ai/claude-agent-sdk at ^0.2.6, and that SDK is the component doing the actual agent execution. When the SDK changes its interface, this app has to follow, and there is no test suite in package.json to catch the breakage automatically. The Electron version is also a moving target, and each bump can invalidate the compiled better-sqlite3 binary, which is why the rebuild script is part of the normal workflow rather than an emergency measure. Budget for periodic rebuilds, not just version bumps. ## Conclusion Adopt it if you already have Claude Code installed and authenticated and you want a windowed view of agent sessions instead of terminal scrollback. Do not adopt it if you need a stable, documented tool: the README is thin, the roadmap section is an empty todo list, and the last push to main was on 2026-03-21, so treat it as a snapshot rather than a maintained product. Before you commit, verify three things: that the release you download matches the platform you are on, that your ~/.claude/settings.json is picked up as the README claims, and that the MIT licence file is actually present in the tree you clone, since the repository listing at the top level does not show one. ## FAQ ### How do I install Open Claude Cowork? Download a release from the project's Releases page, or build from source with Bun or Node.js 22 or newer. The source path is git clone, then bun install, then bun run dev for development mode or one of the dist: scripts for a packaged binary. ### What is Open Claude Cowork? It is an open-source desktop AI assistant, package name agent-cowork, that wraps the Claude Agent SDK in an Electron application for programming, file management and other describable tasks. The README positions it as an alternative to Claude Cowork with a graphical interface instead of a terminal. ### How do I use Open Claude Cowork? The README says you create tasks and choose execution paths, without needing to learn the Claude Agent SDK. It reuses your existing ~/.claude/settings.json, and the README's example shows a task that organizes a local folder. ### Does Open Claude Cowork run on Windows? package.json includes a dist:win script that builds a Windows x64 binary, so the build path exists. The README's platform badge, however, lists only macOS and Linux, so the documentation and the build scripts disagree. ## Sources - [DevAgentForge/Open-Claude-Cowork on GitHub](https://github.com/DevAgentForge/Open-Claude-Cowork) - [Issues](https://github.com/DevAgentForge/Open-Claude-Cowork/issues) - [README](https://github.com/DevAgentForge/Open-Claude-Cowork/blob/main/README.md) - [Releases](https://github.com/DevAgentForge/Open-Claude-Cowork/releases) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/devagentforge-open-claude-cowork