21st Aug 2026

Let’s be honest about where most professional AI use actually lives. You’ve got ChatGPT or Claude open in a tab, you type a prompt, it hands you back text. You draft the email, it summarises the doc, you brainstorm, it throws out an idea. Useful. But it’s still a chatbot. It talks about the work. It doesn’t do the work.

There’s a gap there, and I’ve felt it in my own lab. An AI that can describe how to organise a folder, and an AI that actually goes and organises it, are two different animals.

That gap is what OpenWork is built to close. It’s a free, open-source framework from Different AI — a sovereign, model-agnostic alternative to Anthropic’s Claude Cowork. The pitch isn’t “another chatbot”. It’s an “AI coworker” that lives on your machine and runs the multi-step tasks you used to do by hand.

Here’s why I think it’s worth a professional’s time, a sceptical eye, and a little bit of trust.

From “write me a script” to “go do this”

The biggest shift isn’t technical. It’s that you stop prompting and start delegating.

A normal LLM is great at generating text and stops there. Ask it to “write a script to organise my files” and it writes a script. You’re still the one who has to run it, check it, and fix it when it breaks.

An agent like OpenWork is an execution layer. It sits on top of your filesystem and your tools. So instead of asking for code, you hand it an outcome: “Organise every PDF in Downloads by year and drop them into Archive.” And it goes and does it — fires up a session, walks the folders, moves the files, and checks its own work.

That’s the whole shift. You’re no longer the typist, checking each line of output. You’re the manager, defining the result and reviewing it. That’s where the actual time savings live, the hours that quietly come back to you when you stop doing the clicking.

Your data stays at home (almost)

This is the part that lands in a room full of professionals.

In most firms – legal, healthcare, finance in particular – the reason you can’t just “use the AI” is compliance. Hand a client’s file to a third-party cloud and you’re heading straight for GDPR, HIPAA, and whatever your own policy says. It’s usually a firm no.

OpenWork’s answer is to run inside your trust perimeter. It’s open source, and it runs on hardware you own – which might be a laptop, or your own private cloud. Which means the files it works on don’t have to leave your network. You decide what goes out, and by default it’s nothing.

That’s how a regulated org gets to use a frontier model without the compliance team putting their head on a table.

I’ll add one honest asterisk, because I’m not going to sell you something and hide the catch. “Local” is true for the app and the data it touches. A few of the surrounding features – the shared capability catalogue, the bit that runs the enterprise plane – go through the vendor’s hosted endpoint. Your client files are not what’s going over that pipe. But it’s a pipe you should know about, and I went into it in more detail in an earlier post (OpenWork: The Claude Cowork Alternative) – read that one before you trust it with your files.

Pick your model, not your vendor

Almost every “enterprise AI strategy” out there is really a loyalty programme for one vendor. Use ChatGPT and you’re married to OpenAI’s pricing and roadmap. Use Claude and you’re married to Anthropic’s. Day one.

OpenWork is model-agnostic. It speaks the Model Context Protocol (more on that below) and plugs into well over 50 LLM providers, cloud and local:

• Cloud: Anthropic, OpenAI, AWS Bedrock, Google Vertex AI.
• Local: Ollama, if you want fully offline, fully private inference on your own GPU. (My own preference is llama.cpp as readers of my blog here would have guessed :-P)

That “bring your own key” flexibility is a big deal, and it’s where the sceptic in me starts doing the maths. Route the hard reasoning to a big expensive model, and the quick data-parsing to a small cheap local one. You’re no longer paying premium rates for a query a small model could have handled. And if one vendor’s API has a bad afternoon, you’re not the one holding the phone.

(If you run a homelab — and if you’re reading this site, a few of you probably do — “local via Ollama” isn’t a slogan. It’s a box in the rack that’s already there and already paid for.)

The connector that makes it useful: MCP

Here’s the part that actually makes the agent useful in a business, and it’s the piece most people skip in the demo.

OpenWork has native support for the Model Context Protocol (MCP) — an open standard Anthropic put out that lets a model securely talk to the tools in your business. Think of it as a universal adapter, the way a USB-C port is to a laptop.

With MCP, an OpenWork agent can discover and use:

• The databases you already have (Salesforce, HubSpot, an old SQL box in the corner).
• Proprietary APIs you’ve built up over the years.
• Local scripts and runbooks that live in your code repo and get used by exactly no one.

So a task looks like: “Pull last month’s sales from Salesforce, find the trend, and make a PowerPoint for the Monday meeting.” It grabs the data, crunches it, formats the deck. One pipeline, one instruction. That’s the bridge between “AI is smart” and “AI is actually in our infrastructure”.

The part for the IT people: OpenWork Den

If you’re a homelabber, skip this section. If you’re in IT, this is the one you’ll actually be asked about.

OpenWork is more than a desktop app. It’s a framework, and on top of it sits “Den”, the enterprise control plane. What it gives you:

• Provision AI inference across the whole company, not per-nerd.
• Manage who can access what.
• Curate an internal marketplace of pre-approved tools and plugins, so a junior doesn’t bolt on a random MCP server and open a door.

Put differently: employees get the power of an autonomous assistant, and the organisation keeps the keys. That’s the deal IT has been trying to strike with AI for years — the “democratise it, but don’t let it out the back door” deal. OpenWork argues it can actually strike that deal.

Whether you need Den depends on your size. A five-person shop probably doesn’t. A company with a compliance function probably does. (My own read, from the other side of the fence, is “impressive, and completely unnecessary” — the honest take for a solo operator.)

One build, many users

This is the one that makes cross-functional teams sit up.

The normal failure mode is this: an engineer writes a tidy Python script to automate a report. It’s great. It lives in a git repo. And the marketing manager who actually needs the report has, essentially, never opened a git repo in her life. The automation exists, and it’s useless to the people who need it.

OpenWork lets that engineer wrap the script into an MCP tool. Now the marketing manager triggers the whole pipeline with a plain sentence — in the OpenWork GUI, from Slack, or from Telegram. No code, no git, no “please file a ticket with engineering”.

That’s a silo, demolished. The technical team builds it once, and the non-technical team uses it the same day. The busywork doesn’t get cleverer. It just gets automated, for the whole org.

If you’re going to try it

OpenWork is free, open source, and community-built. If you want to move past the chatbot stage, here’s the path I’d take, in roughly that order:

1. Read the code. The repo is different-ai/openwork. I’d actually read it before I let a tool move my files. It’s open source, so you can.
2. Start small and boring. Install the desktop app, and give it a task that’s low-risk and reversible on your own machine – tidying a folder, summarising a local doc. If it blows up, nothing you care about is in the blast radius.
3. Watch it work. The GUI has a “Visual Safety Dashboard” that shows you the plan ahead of each step, and it pauses for your permission on the sensitive ones. That pause is the feature. That’s you keeping the veto.
4. Only then, think about scale. If you’re in IT and you want this across the firm, go look at Den and how it would deploy across your network. Do that with your security hat on, not your excited hat on.

The point

Learning OpenWork isn’t really learning a tool. It’s learning a different job description for the work. The AI is going from a passive assistant you talk to, to an active collaborator you delegate to. The professionals who get comfortable handing a multi-step workflow to an agent – and then reviewing the result like a manager – will have a real edge over the ones still copy-pasting between a prompt and a text box.

The future of work isn’t about writing better prompts. It’s about building better automations. And OpenWork makes the first step into that free, open, and something you can actually read the source for.

That said, I’m still running it in my own lab, watching it like a hawk, and all looking good so far. I’ll keep you posted when that changes.