There are two ways to get AI help with your work. Bring the AI into your workspace, or bring your workspace to the AI. Custom agents do the first. MCP does the second.
Atwork runs an MCP server, which means an assistant like Claude can connect to your workspace directly — read your sheets, create tasks, post comments — without you copying anything into the chat. MCP is an open standard for connecting AI assistants to external systems; that is the last you need to think about the protocol.
What does connecting your workspace to Claude let you do?
It lets you ask questions about your real data, in the assistant you are already talking to, and have the answers act on your workspace rather than just describe it.
In practice that means things like:
- Read a sheet and answer a question about it. Which deals have not moved in three weeks. How many open bugs are unassigned. What did we ship last month.
- Create tasks from a conversation. You have just talked through a plan with the assistant; it writes the resulting tasks into the right sheet rather than handing you a list to retype.
- Post a comment on an item, so the outcome of a discussion lands where your team will see it.
- Read a dashboard or a page and summarise what it shows.
The common thread is the elimination of copy-paste. Not a small thing: copy-paste is where context gets lost, where data leaves your workspace, and where the result fails to make it back in.
How is this different from a custom agent?
They solve adjacent problems and the choice between them is mostly about where you are sitting.
A custom agent lives in your workspace. Your whole team uses it, it appears in assignee dropdowns, it picks up tasks and hands them back. It is configured once by an administrator and shared. Use it for recurring work that belongs to the team.
An MCP connection lives in your assistant. It is yours, it follows your chat, and it is ad hoc by nature. Use it for the question you have right now, especially when the work spans your workspace and things outside it — a document you are drafting, a codebase, a spreadsheet on your machine.
The honest summary: agents are for repeatable team processes, MCP is for individual exploratory work. Most people who use both end up reaching for MCP more often and getting more durable value from agents.
There is a third option people forget, which is that a lot of recurring work needs no AI at all — see agent or automation for how to tell.
What can the assistant see through the connection?
What you can see. The connection authenticates as you, so the assistant reaches exactly the spaces, sheets and items your own account reaches — no more, and nothing that belongs to colleagues you cannot already see.
The same principle governs in-app agents, which always act as a real person rather than as a privileged service account — the permission model is worth reading in full if you are the person who has to sign off on this. There is no shared key across your organisation and no bypass of your workspace permissions.
What does change is that data crosses into your AI assistant's context when you ask a question that requires it. That is the actual consideration to weigh — not unauthorised access, but deliberate sharing with a model provider. If your organisation has a policy about which data may go to which AI vendor, that policy applies here, and the connection does not override it.
What is the setup like?
Atwork exposes its MCP server at a fixed address, and you add it to your assistant as a connector, then sign in with your Atwork account to authorise it. Sign-in happens once; after that the connection persists and your assistant can use it in any conversation.
The exact clicks depend on which assistant you are using and change as those products evolve, so we are keeping this section deliberately short rather than publishing steps that will be stale in a month. The shape is always the same: add the connector by URL, authorise with your workspace account, confirm the tools appear. The Atwork MCP page has the current steps for each assistant. The consent screen lists every action the connection asks for and you allow it as one set; connecting Atwork to your AI assistant explains what that grants and what a connected assistant is never offered.
One behaviour worth knowing in advance: if the connection has been idle for a long stretch, you may need to sign in again. This is ordinary session expiry rather than a fault.
What is it good at, and what is it bad at?
After using this daily, the pattern is fairly consistent.
Good at:
- Questions that need reading across many rows, where you would otherwise build a view or scroll
- Turning a discussion into structured records without retyping
- Cross-referencing your workspace against something outside it
- One-off analysis that does not deserve a permanent dashboard
Bad at:
- Bulk edits. Ask it to update forty rows and you are trusting forty judgment calls you have not reviewed. Use a view and edit them yourself, or an automation.
- Anything you need to be exactly reproducible. The same question can produce differently-shaped answers on different days. If you need the same number every Monday, build a dashboard.
- Precision addressing when you are vague. Ask it to update the onboarding task and it may pick a near-match. Name the sheet and the item.
That last one is the mistake most worth internalising: be specific about which item you mean. Ambiguity in, wrong row out.
When should you use this instead of just opening Atwork?
When the answer requires synthesis rather than lookup. Opening the workspace is faster for anything you can find with a filter — and you should keep doing that. The connection earns its place when the question crosses several sheets, or when you are already mid-conversation with an assistant and breaking out to a browser would lose the thread.
It also earns its place for the question you cannot quite express as a filter. What is going wrong with this project is not a query, but it is a reasonable thing to ask something that can read the project.
Try it on a real question
Connect it, then ask about something you actually need to know this week — not a test question. The value shows up immediately or it does not show up at all, and a real question tells you which within about a minute.
Need a workspace first? Start free, then point your assistant's connector settings at it.