OpenWork Setup: Open-Source Claude Cowork in 15 Minutes

OpenWork is a free desktop app that hands an AI agent a folder on your own machine and lets it do real work there, and today it is the fastest-climbing repository on GitHub trending, up roughly 915 stars in a day on 18,900 total. Budget about fifteen minutes for setup, most of it spent pulling a 169 MB Electron build, and know that you can skip the app entirely if you would rather bolt its MCP server onto Claude Code, Codex or OpenCode.

• v0.18.12 shipped on 30 July with installers for Windows, macOS and Linux across x64 and arm64.
• The desktop app is optional. One remote MCP URL exposes the same skills, plugins and connected services inside whatever agent you already run.
• Licensing is open core: MIT across most of the tree, a Fair Source license for everything under /ee.
• It runs on opencode underneath, so you supply your own model keys instead of buying a seat from a single vendor.

What is OpenWork and why is it trending?

Different AI open sourced OpenWork as an alternative to Claude Cowork and Codex, and the repository is only about six months old. It was created on 14 January 2026 and has already collected 18,947 stars, 1,924 forks and 67 contributors. Different AI is listed as a Y Combinator company, and a Show HN post put the project in front of the usual early-adopter crowd, which is roughly where the star curve turned vertical.

Relatedego lite Setup: Let Claude Code Drive Your Real Browser

The pitch that seems to be landing is not the desktop app. Plenty of projects ship an Electron shell around an agent. What OpenWork adds is a distribution layer: you configure a skill, an MCP connection or a Google Workspace hookup once, and then reach it from Claude Code on your laptop, Codex on a work machine and a teammate's install, without re-doing any of it. The agent runtime underneath is opencode, so model choice stays yours.

Anthropic's Cowork and OpenAI's Codex both solve the first half of this and keep the second half inside their own product. OpenWork is betting that teams want the capability layer to be portable.

How do you install OpenWork on macOS and Linux?

There are two routes. The scripted one installs a small bootstrap CLI that then fetches the right desktop build for your platform. The official instructions are explicit that you should not pipe the installer straight into a shell, so download it, read it, then run it.

# download, inspect, then run. Requires Node.js 20 or newer
$ curl -fsSLo /tmp/openwork-install.sh https://openworklabs.com/install.sh
$ less /tmp/openwork-install.sh
$ sh /tmp/openwork-install.sh

The script drops a dependency-free Node file into $HOME/.local/bin and needs no npm, npx or admin rights. Confirm it landed:

$ openwork-bootstrap doctor --json

Now pull the desktop application itself from the published manifest, which resolves your OS and architecture for you:

$ openwork-bootstrap install app \
    --manifest https://openworklabs.com/install-manifest.json \
    --json

# verify the app half of the install
$ openwork-bootstrap doctor --app --json

Create your first workspace. The --prepare-desktop flag wires it into the app so it opens ready to run, and --owner-email is optional convenience that pre-fills the claim page later:

$ openwork-bootstrap cloud bootstrap-workspace \
    --base-url https://api.openworklabs.com \
    --workspace-name "my-workspace" \
    --skill-name "First OpenWork Skill" \
    --claim-roles owner \
    --prepare-desktop \
    --json

On macOS, launch it with open -a OpenWork. Linux users who prefer to skip the CLI can grab openwork-linux-x86_64-0.18.12.AppImage at 159.3 MB or the 214.2 MB x64 tarball from the Releases page. An arm64 AppImage and tarball are published alongside them.

How do you install OpenWork on Windows?

The bootstrap script is POSIX shell, so on Windows take the installer directly from the v0.18.12 release. The file you want is openwork-win-x64-0.18.12.exe at 168.9 MB, or openwork-win-arm64-0.18.12.exe at 159.8 MB on Snapdragon hardware. Three variants exist per platform and the difference matters:

• openwork-* is the plain public build.
• openwork-cloud-* requires sign in without enterprise activation.
• openwork-enterprise-* keeps the standard app identity but targets managed deployments.

Unless someone handed you a licence, take the plain build. One caveat straight from the release notes: the Windows installer is temporarily unsigned while production code signing is finalised, so SmartScreen will interrupt you and you will have to click through it. If your organisation blocks unsigned installers, wait for a later release rather than fighting policy.

How do you wire OpenWork into an agent you already use?

This is the part worth doing even if you never open the desktop app. The MCP server exposes exactly two tools, search_capabilities to find what is available to you and execute_capability to run it. Adding it opens a browser so you can sign in and pick your organisation.

# Claude Code
$ claude mcp add --transport http openwork https://api.openworklabs.com/mcp/agent

# Codex
$ codex mcp add openwork --url https://api.openworklabs.com/mcp/agent
$ codex mcp login openwork

OpenCode wants it declared in opencode.json instead:

RelatedStrix Setup: Run an AI Penetration Tester on Your Code

{
  "mcp": {
    "openwork": {
      "type": "remote",
      "enabled": true,
      "url": "https://api.openworklabs.com/mcp/agent",
      "oauth": {}
    }
  }
}

Any other MCP client points at the same remote URL. Restart the client afterwards, and do not assume it worked until the restarted agent actually lists both tools.

What are the gotchas before you rely on it?

Start with the licence, because "open source alternative" is doing some work in that description. The bulk of the tree is MIT, but everything under /ee carries a Fair Source licence, and that directory is where the commercial control plane lives. This is standard open core and it is disclosed plainly in the LICENSE file, but it means the org management surface is not something you can freely fork and resell.

Second, understand what the provisional workspace actually is. The bootstrap flow deliberately creates a workspace with no authenticated owner so you are not forced through an account signup first. The setup documentation says outright that the optional owner email is a convenience for pre-filling the claim page, not an account and not a security boundary. Ownership only becomes real when a human claims the workspace through a browser link. Do not put anything sensitive in an unclaimed workspace.

Third, the pace. There are 414 open issues on a project that shipped its last commit yesterday and is cutting releases at a rate that already carries it to v0.18.12 in six months. That velocity is why the features look good, and it is also why you should not build a team process on any specific behaviour this week.

Smaller ones: the bootstrap CLI requires Node 20 or newer and installs to $HOME/.local/bin, which is not on PATH on every distribution. Teammate invites do not go out until the workspace is claimed, because there is no authenticated sender until then. And the desktop builds are Electron, so roughly 160 to 215 MB per download is the floor.

Our take

Install the MCP server first and treat the desktop app as optional. That inverts the README's ordering, but it matches where the value is: a skill you configure once and can reach from Claude Code, Codex and a colleague's laptop is worth more than another Electron window competing for space with the editor you already have open. Fifteen minutes gets you the MCP endpoint wired into whatever agent you use daily, and you can decide about the 169 MB download afterward.

The scepticism to hold onto is about durability rather than quality. A six-month-old project with 400 open issues and a control plane behind a Fair Source licence is not a foundation, it is a bet. The bet is reasonable, since portable capabilities across agents is a real gap and nobody with a vendor incentive is going to fill it. Just keep your skills in a form you could re-create elsewhere.

Original analysis by GenZTech. Tool documentation: different-ai/openwork on GitHub.