Three things I built in Claude Cowork in a matter of hours.

Claude Cowork launched in January 2026 as an agentic mode inside the Claude Desktop app. The idea is straightforward: instead of describing a task to Claude in chat and then going off to do it yourself, Cowork can actually do it. You give it access to a folder on your machine, define the goal, and it handles the steps — reading files, creating files, organising things, pulling in external data through MCP connectors.

It's built on the same architecture as Claude Code, but without the terminal. The target user is knowledge workers rather than developers. You point it at a folder, not a codebase.

It went generally available across all paid plans in April 2026 — macOS and Windows both supported. Before that it was a research preview. I got access during the preview period and have built three automations with it. Here's what each one does and what I noticed.

Folder cleanup on a daily schedule

My downloads folder and project directories accumulate fast. Screenshots, exports, one-off files that don't have a home. I built a scheduled automation that runs each day and sorts everything — moves files to the right folders, applies consistent naming, clears out the obvious junk.

It's not sophisticated. But it runs without me thinking about it, and the cumulative effect of having a consistently organised file system is bigger than I expected. I find things faster. I start work from a cleaner state. It's the kind of maintenance task that's easy to skip until it isn't.

Daily AI news digest

Every morning, before I open anything else, a scheduled automation runs a sweep of the latest AI news — new tools, model releases, research, anything worth knowing — and produces a written digest. Each item gets a brief summary. It also pulls YouTube recommendations on AI topics with a short description of what each video covers.

I was previously doing this manually, badly, and inconsistently. Now it's just there.

The useful constraint it introduced is that I read the digest first rather than grazing across six tabs. The format forces prioritisation — a short summary either earns more of my time or it doesn't. That's changed how I stay across the space more than the time saving has.

Monthly Lovable financial reporting

I run several Lovable projects — including production builds for clients — and was doing a manual reconciliation each month. Pulling credit usage, categorising spend by project, working out where the costs were going. Tedious and easy to let slide.

I built a Cowork automation with a custom MCP connector that pulls directly from Lovable. It generates a monthly breakdown: credits used per project, spend by category, month-on-month comparison. Runs on schedule, lands in the same place every month.

Cowork's MCP support is what makes this worth doing at the right level of fidelity. Anthropic released a set of open-source plugins for Cowork in January built on the same MCP protocol that Claude Code uses — so the connector pattern is consistent whether you're working in a terminal or the desktop app. In this case, pulling structured data directly rather than exporting a CSV manually means the report is always current without me touching it.

A few observations

The pattern across all three is the same: tasks I knew I should be doing regularly but had a high enough setup cost that I'd let them slide. Cowork lowered that cost enough that building the automation became the obvious move.

The scheduled execution is underrated. An AI tool you have to remember to open is a different product to one that runs at 7am and has something ready for you. The latter changes your habits; the former competes with them.

One thing worth understanding about how Cowork works: it has access only to the folders you explicitly give it. Everything runs locally. Anthropic doesn't train on your files. For anyone building automations that touch client or financial data, that matters.

The MCP connector approach for the Lovable reporting is something I want to extend. There are other data sources I'm reconciling manually that could plug into the same pattern — same structure, different connector. That's probably where most of the value is for business use: not one-off tasks but recurring integrations that just run.

If you want to talk through how AI could optimise your workflows, get in touch.