OpenWork is a free, open-source desktop app and MCP that shares AI skills and connectors across a whole team, the same job Anthropic's Claude Cowork does inside its paid Team tier.
Get a whole team, an ops lead on Claude Code, a marketer on Cursor, a contractor on Codex, working off the same set of AI skills and connectors, and Anthropic has exactly one paid way to do that today: put everyone on a Claude Team seat. That runs $20 to $25 a seat a month for the standard tier and $100 to $125 a seat for premium, according to Claude's own pricing page. That fee isn't really paying for the model, Pro already includes Claude Cowork at $17 to $20 a month for one person. It's paying for the admin layer: the shared connectors, the org-wide skill library, the thing that turns Cowork from a solo laptop trick into something a team can run together. OpenWork, a desktop app that picked up 628 stars in a single day on GitHub's trending list this week and now sits at 18,520 stars overall, ships an open-source version of that same admin layer, and it works across Claude Code, Cursor, and Codex instead of locking a team into one vendor's app.
OpenWork's README describes the project plainly as "the open-source alternative to Claude Cowork and Codex," built on top of the opencode framework. The mechanism is one MCP server. Add it to Claude Code, Cursor, Codex, or any other MCP-compatible agent, and it exposes two tools: one that searches what capabilities are available, and one that runs them. Skills, MCP connections, and Google Workspace or Microsoft 365 integrations configured once now follow into whichever agent is open, instead of getting rebuilt inside each tool's own settings. The desktop app itself is optional; the whole thing can be driven from a coding agent someone already has installed.
The piece that actually maps to what a Claude Team seat charges for is OpenWork Den, the project's admin control plane. It lets an organization publish skills and plugins to a shared internal marketplace, assign them to a team or a single person, choose which model providers each group can use, and set desktop policies restricting local model access, the same category of control Anthropic's own pricing page lists under "admin controls for remote and local connectors" and "organization wide skills deployment." OpenWork can also import Anthropic-compatible plugins directly, so a team that already built skills for Claude isn't starting from zero.
For a marketing ops lead or a small agency running client work across three or four AI tools because different people on the team prefer different agents, this closes a real gap. Teaching an assistant to draft a specific client report format, or wiring up a connector to the CRM, is currently work done once per tool, per person. A shared skill library any teammate's agent can pull from turns that from a repeated setup tax into a one-time investment, without requiring the whole team to standardize on Claude specifically to get it. That is the part of the Team tier's value that has nothing to do with model quality, and it's the part OpenWork is aimed squarely at replacing, at no license cost, for whatever mix of agents a team already runs.
Here's the honest part. Most of what's active on the releases page right now isn't the one stable build tagged today, it's Canary and Experimental prerelease channels, one of them unlocked by tapping a hidden developer-mode switch five times in eight seconds, with the Windows installer explicitly flagged as temporarily unsigned while production code signing gets finished. Running OpenWork Den yourself also means being the one who keeps the control plane patched and online, work Anthropic's Team and Enterprise plans bundle with SSO, SCIM, and audit logs on infrastructure Anthropic operates. And a direct fetch of the repo's license file returned a 404 during this pass, so take the "open-source" framing from the README at face value rather than a verified license grant until someone on the team actually opens the file. This is a fast-moving project, not a finished product to install once and forget.
Anthropic didn't build a $20-to-$25-a-seat fee around the model. It built it around the admin panel that lets a team share the model's setup, and that's the layer open source just cloned first.