OpenWork AI Explained: What It Is, How To Set It Up, Every Feature, And The Best Models To Run In It

OpenWork AI is a free, open-source desktop app for macOS, Windows and Linux where AI agents do real work on your own files, and it is built on OpenCode as an alternative to Claude Cowork.

Different AI, Inc. makes it, and its founder and CEO is Ben Shafii.

The desktop app is MIT-licensed, and you can use it locally without paying for it or creating an account.

You bring your own model, which can be a ChatGPT sign-in, an API key or a local model.

In this post I'll explain what OpenWork is, what it costs, how to install it, how every main feature works and where its limits are.

After that, I'll cover the question that fits this site, which is which model you should actually run inside it.

I checked every OpenWork fact against its own GitHub repo, docs and pricing page on 10 October 2026.

What OpenWork is

OpenWork is a workspace where you pick a folder, pick a model and hand an AI agent a job.

The agent reads your files, runs tools, calls MCP servers and drives a built-in browser.

The README calls it "the free, open-source alternative to Claude Cowork and Codex".

It runs on OpenCode, the open-source agent, which is why my OpenCode walkthrough sits at the top of this page.

Ben Shafii wrote in the launch post that he built it to bring terminal agent workflows into an interface non-technical people could use.

Which OpenWork this page covers

More than one project has used this name, so I want to be precise.

This page covers OpenWork from Different AI at openworklabs.com.

In January 2026 I tested a different app called Openwork on camera, and that app was distributed from accomplish.ai.

That project was later renamed, and its GitHub repo now says it is no longer supported.

There is also an npm package called openwork from LangChain, and OpenWork's docs warn you not to run npx openwork.

I haven't done a full on-camera test of today's OpenWork yet, so the product facts here come from its own pages.

What OpenWork costs

OpenWork also sells two add-ons on paid plans.

OpenWork Cloud Computer costs $50 per member per month, and OpenWork models cost $10 per user per month.

The licence is split by folder.

Everything outside the ee/ folder is MIT, and the ee/ folder holds the team control plane under a source-available licence.

That second licence is free for organisations of up to 5 users, and each release converts to MIT after two years.

The important point for this site is that the free app does not include a model.

You pay for tokens through the provider you connect, unless you run a local model.

How to install and set up OpenWork

OpenWork's docs give the first run as four steps, and none of them needs an account.

First, download it from openworklabs.com/download, which has builds for macOS, Windows and Linux.

Second, pick a folder, because OpenWork only works in folders you authorise.

Third, choose a model under Settings › AI Providers.

Fourth, run a task, and the docs suggest "Summarize this folder."

On a Mac you can also install it with Homebrew.

brew install --cask openwork

The Homebrew cask needs macOS 12 or newer, and it can sit one version behind the newest release.

You can also grab any build from the GitHub releases page.

How OpenWork connects to models

This is the part that decides both your cost and your results.

For Ollama, you wait for the status to say "Ollama running", pick or pull a model, and click "Add to workspace".

OpenWork reads whether each model supports images and which thinking levels it offers.

The docs say a local model needs no API key, no OpenWork Cloud account and no internet once it has downloaded.

Every main OpenWork feature

Skills use the same SKILL.md format as Claude.

To create one, you finish a job in chat and type, "Turn what we just did into a reusable skill for me."

The built-in browser asks once per thread before it navigates and reads pages.

Clicks, typing and keyboard input ask every time, because they can change data on a site you're signed in to.

The docs say the browser doesn't get access to your regular Chrome profile.

The workbook tools keep numbers as numbers, but they don't make charts, pivot tables or macros.

The MCP Gateway is a single command in Claude Code.

claude mcp add --transport http openwork https://api.openworklabs.com/mcp/agent

OpenWork limits

OpenWork is on version 0.18, and the repo showed about 600 open issues and pull requests when I checked.

Its roadmap marks scheduled tasks and sandbox workspaces as Partial.

Its docs say cross-chat memory is not a separate long-term memory database, and nothing is recalled automatically.

Managed connectors, Library installs, automations and the MCP Gateway all need an OpenWork Cloud sign-in.

The docs say there is no result cache, so rerunning the same task costs tokens again.

If you migrate from Claude Cowork, conversation history and credentials don't come with you.

OpenWork vs Claude Cowork, Hermes Desktop and my Agent OS

Claude Cowork is the more finished product today.

Anthropic's help page says Cowork is available on paid plans across desktop, web and mobile.

OpenWork's own comparison marks OpenWork as partial for memory across chats, sandboxing, automations and run history.

OpenWork is ahead on model choice, Linux support, price and self-hosting.

Hermes Desktop is the official app for Hermes Agent, and my Agent OS runs a crew of agents under one shared memory.

The difference that matters most on this site is the first one, because OpenWork lets you choose the brain.

OpenWork doesn't ship its own model

OpenWork is an app, so it isn't on GoldieBench, and it doesn't come with a model of its own.

OpenWork does sell managed inference as a paid add-on, but the free app expects you to connect something.

That makes the model you pick the single biggest factor in how good your results are.

OpenWork's docs make the same point about local models.

They say small local models often chat well but struggle to call tools reliably.

They also say to try a larger model if a task stalls, and to keep a cloud model connected as a fallback.

How the models you'd run in OpenWork score on GoldieBench

GoldieBench gives every model the same one-shot prompt to build a game, a page, a simulation or a visual, then scores the rendered result from 0 to 10.

Here are the models people are most likely to connect to OpenWork, with their live averages.

Those numbers measure one-shot building, so they tell you how strong each brain is at producing working output.

They don't measure tool calling inside OpenWork, so I can't promise the same order for long agent tasks.

DeepSeek V4 Pro is the open model OpenWork's own cost calculator uses as its default example, but it doesn't have a score on the board yet, so I won't guess one.

What the bench means for your OpenWork setup

The gap between local models is the lesson here.

A 27B local model scores 7.14 on the board, which sits right next to cloud models like Gemini 3.6 Flash and Claude Sonnet 5.

A 12B local model scores 3.98, which is a very different experience.

That matches the warning in OpenWork's docs about small local models.

So if you want OpenWork to be free per task, you need enough memory to run a bigger local model.

The local models board shows which free models are worth running on your own machine.

If you'd rather pay for quality, the top cloud models on the board are a clear step up.

My rule is to use a cheap or local model for reading and summarising, and a top model for anything that builds something.

OpenWork lets you switch models from the picker under the message box, so you can do that inside one workspace.

For a deeper ranking, read my best LLMs for Hermes Agent post, because the same logic applies to any agent.

The best Hermes Agent model post and the best Hermes Agent alternative post cover the nearby questions.

Should you use OpenWork?

I'd try OpenWork if you want Cowork-style file work and you want to choose the model yourself.

I'd stay with Claude Cowork if you already pay for Claude and you want the most polished experience.

Either way, the model matters more than the app, so pick it with real scores and not with hype.

Also On Our Network

• Agent Operatorshow to install OpenWork and run your first task
• AI Income Deskseven real jobs a business can give OpenWork
• AI Tool Verdictevery OpenWork feature explained and scored
• agentos.guidean OpenWork quick-start for your Agent OS
• agentos.guideOpenCode and Hermes as a free AI dev team, on agentos.guide

Install the free app, connect a model the bench says is strong enough, and you'll get far more out of OpenWork AI.