Open Cowork: What the Open-Source Claude Desktop Reveals About AI Agent Tooling
Claude Code is one of the most capable AI coding agents available. It can read your repo, follow your conventions, plug into MCP servers, run custom skills, and ship production code. Most developers know this.
Most non-developers have no idea it exists.
That's not a complaint — it's a design constraint. Claude Code lives in a terminal. It requires an API key, a config file, and enough comfort with a command line to type claude and understand what happens next. That's a real barrier if you're a product manager, a designer, or someone who just wants to ask an AI to organize their files without opening iTerm.
Open Cowork ships what Claude Code can't: a one-click desktop installer for Windows and macOS. No terminal. No config files. Download the DMG, paste your API key, pick a workspace folder, and you're talking to Claude.
The interesting part isn't the GUI wrapper. It's what they built underneath it — and that they gave the whole thing away under MIT.
What it ships
Open Cowork is an Electron app built on the Claude Code Agent SDK. The architecture is straightforward:
- Main process (Node.js): Handles the agent runner, session management, tool execution, and MCP connectors
- Renderer (React + Tailwind): Chat UI, trace panel, settings, sidebar
- Sandbox layer: Routes all Bash commands to an isolated VM — WSL2 on Windows, Lima on macOS — so the agent can't touch your host filesystem outside the workspace
That last piece matters. When you give an AI agent file access on your machine, the scariest failure mode isn't "it writes bad code" — it's "it deletes something it shouldn't." The VM isolation is the engineering decision that makes it safe to actually use.
The Skills system is the other thing worth understanding. Skills are pre-built Claude workflow definitions — essentially Claude Code skill files — for document generation: PPTX, DOCX, XLSX, PDF. You point Claude at a CSV or a folder of notes and say "make me a deck," and the skill handles the extraction, formatting, and output. You can add your own.
That pattern — giving Claude reusable workflows rather than hoping it rediscovers the same approach every session — is exactly what OTF kits do with .claude/skills/. The concept translates directly.
The five demos
The repo ships five demo videos that are worth watching before forming an opinion. They're not toy examples:
Folder organization & cleanup — Claude reads a messy directory, proposes a structure, and reorganizes it. The trace panel shows every tool call in real time.
Generate PPT from files — Drop files into the chat, ask for a presentation. The PPTX skill kicks in and produces a formatted deck.
Generate XLSX spreadsheets — Same pattern for Excel. Useful for anyone who works with data but doesn't want to write formulas.
GUI operation — Claude controls desktop GUI applications via computer use. They recommend Gemini 2.5 Pro for this; it reads screen state better than most models on GUI understanding tasks.
Remote control via Feishu — Send a message in Feishu (Lark), Claude executes the task and replies. Same idea works with Slack. This is the asynchronous agent pattern — you don't have to be at your computer for it to work.
These aren't polished marketing clips. They're screen recordings of real workflows. The rough edges make them more credible.
Why open source changes the economics
Most AI agent GUI products are hosted SaaS — you pay per seat, per generation, or per output. The moment the vendor decides to pivot or shut down, your workflow dies with it.
Open Cowork is MIT. You download the source, you build the binary, you own the binary. If the project stops shipping updates tomorrow, nothing breaks for you. If you want to add a skill that generates SOW documents for your consulting workflow, you fork the repo, add .claude/skills/sow/, and ship your own build. No approval process, no waiting for a feature request to land.
That's the open-source multiplier — not just that it's free, but that the architecture compounds. Every skill someone builds can be shared. Every integration someone adds becomes a PR. The MCP connector for Notion that one contributor wrote is now everyone's.
Compare this to a proprietary equivalent. The vendor decides what integrations exist, when they ship, and how much they cost. The ceiling is whatever the product team can prioritize. Open Cowork's ceiling is whatever the community can build.
The model support story
One architectural decision that's easy to overlook: Open Cowork supports any OpenAI-compatible API endpoint. That means Anthropic directly, OpenRouter, or Chinese models like GLM-4, MiniMax, and Kimi.
This matters for international users, for cost optimization, and for anyone who wants to switch models based on task type. GUI automation works better with Gemini 2.5 Pro. Code generation works better with Claude Sonnet. You pick the right model for the job, not the one the vendor decided to support.
What this means for OTF users
At OTF, we make the same bet Open Cowork is making — that AI agents are most powerful when you give them context. We ship CLAUDE.md, .cursorrules, 40+ tested prompts, MCP hookups, and custom skills with every kit because Claude without context is Claude starting from zero every session.
Open Cowork ships the same idea, one level up: Claude without a GUI is Claude locked behind a terminal that most of your users will never open.
If you're building a product where non-technical users need to interact with AI agents, the Open Cowork architecture is worth studying. Electron + Agent SDK + VM sandbox + Skills is a proven stack. The source is right there.
If you're building something where technical users need to ship production code fast, that's where OTF kits live — same philosophy, different altitude.
Star the repo: github.com/OpenCoworkAI/open-cowork. The Discord is active and the team ships fast.