Connectors and automation in Cowork

In part 1 we looked at what Claude Cowork is. In part 2 we made Claude smarter with skills. But skills only help as long as Claude works with what it can actually see. If Claude cannot reach your calendar, your email or the systems you work in, it will still struggle to deliver end to end. This is where connectors come in, and where Cowork starts to look like an actual work system.

What is a connector?

A connector lets Claude fetch context directly from the tools you already use. Instead of pasting emails, downloading documents or explaining what is in your calendar, Claude goes straight to the source.

Technically this is built on MCP, the Model Context Protocol, an open standard that lets AI tools connect to other systems. You do not need to know anything about MCP to get value from it. Think of it as a USB port into your data.

In short: skills tell Claude how to work, connectors give Claude access to the right context. When it has both, it starts to look like systems building.

Three kinds of connectors

Ready-made connectors. Integrations that work out of the box: Gmail, Google Drive, Google Calendar, GitHub, Slack. Log in, connect, done.

Directory connectors. Third-party integrations built on the same standard: HubSpot, Snowflake, Figma, Notion and more, including Microsoft 365 for everyone who lives in Outlook and OneDrive.

Custom connectors. For internal systems and your own APIs. Add an MCP server and connect Claude to exactly what you need.

What it means in practice

One source. Connect Google Drive and say: "Summarise the documents in this folder and write a status report."

Several sources. Connect calendar and email and build a morning brief. Suddenly you get not just today's meetings, but relevant context from earlier threads and things you should be prepared for.

Automatic workflows. Combine a skill, a couple of connectors and a scheduled task. Every Friday Claude pulls data from your tools, runs the analysis and puts the report where it belongs.

This is the shift: before, you had to feed AI its context. Now Claude can fetch it.

How many connectors do you need?

Probably fewer than you think. Most people do not need ten connectors. They need one or two that remove the most friction. Start where you most often do manual context gathering: email, calendar, documents, project tools or CRM. The connector that saves you the most manual work is where the value shows up first.

Our own daily combination is four: Notion as the knowledge base, Gmail and Google Calendar for the morning brief, and Slack for keeping up without being everywhere. Each one is useful alone. Together, Claude actually starts to understand the working day.

Scheduled tasks

Connectors give access. Scheduled tasks give structure. You can ask Claude for things manually every time, but it is when the same tasks run automatically that it becomes a system. The morning brief that is ready when you open the laptop. The weekly report that appears every Friday. The Slack summary after lunch.

Scheduled tasks use the same skills and connectors you have already set up. You are not building something new. You are letting Cowork run it at the right time.

But remember this

When tasks run on their own, mistakes can too. Build simple controls into your skills: ask Claude to flag uncertain points, double-check numbers, and mark things that should be verified before you use them. And expect a few rounds of iteration before workflows deliver exactly what you want. Cowork is still young. Give it responsibility gradually, like a new colleague.

An example

Say you want a weekly pipeline report. You build a skill that knows how the report should be structured: which numbers matter, how they compare to last week, what tone to use. You connect Claude to the CRM and the calendar. Then you set a scheduled task for Friday at two.

From then on it happens by itself: Claude pulls pipeline data, sees which meetings happened, applies your methodology and puts the finished report in your folder. You build it once. After that, the deliverable just shows up.

Getting started

1. Connect one service. Start with the one you use most.
2. Test on something concrete. "Summarise the latest emails from this person" is enough to see the effect.
3. Combine the connector with a skill. That is where the value comes: not just "Claude can read my system", but "Claude uses it to deliver something useful."

Skills give methodology. Connectors give context. Scheduled tasks give independence. Together it is no longer just a chat. It starts to become a system.

Next: the same idea as a way of working, in An agent in a folder.