We Shipped an MCP Connector That Does Less on Purpose

A connector is the easiest way to ruin a product like this one.

Slaet asks one question on your new tab. Is there something you should be doing right now. If there is, it shows you that one thing. If there isn't, it stays quiet and you get on with your day. The whole arrangement rests on a small assumption: that everything in the list was put there by someone who meant it.

Now attach an assistant. Assistants are good at generating plausible next steps. Ask one to help you think through a project and it will hand back fourteen tasks, all of them reasonable, none of them chosen by you. Give it write access and those fourteen land in your Today. Slaet has no defence against that. Every one of them is a real task as far as it can tell, and the tab that was supposed to show you the one thing that matters is now showing you a machine's brainstorm.

This is the default outcome, not the edge case. Most connectors being shipped this year are built to make exactly that easy, and they advertise it as the point.

What we shipped

The Slaet connector is live at https://mcp.slaet.space/mcp. It speaks MCP, the open standard that lets an assistant talk to an app, so it works with any MCP client. Today you add it by hand as a custom connector in Claude, ChatGPT, or whatever else you use. It is not in anyone's app directory.

There are ten tools.

whats_next, list_tasks, capture_task, complete_task, update_task, list_lists, search_notes, create_note, append_to_note, get_note.

The interesting part of that list is what isn't in it. We spent more time deciding what the connector would refuse to do than building what it does.

The surface is narrow on purpose

There is one create path, and it takes one task. No bulk endpoint, no array parameter, no import. An assistant that wants to file ten things has to decide, ten separate times, that each one deserves a place in someone's day. In practice it stops well before ten. That friction is not a gap we intend to close once we have time. It is the mechanism.

whats_next returns one task, or nothing. The same contract as the tab. The nothing matters as much as the one. An assistant that asks Slaet what you should be doing gets a straight answer, and sometimes the straight answer is that nothing needs you.

There is no delete. Not for tasks, not for notes. It is the only irreversible operation in the set, and nothing we wanted the connector to do requires it. Notes can be created and appended to; existing text is never overwritten or removed. The worst thing a confused model can do to a note is leave a paragraph at the bottom of it that you have to read and dismiss.

Reads are capped, and they say so when they truncate. A model that silently receives the first slice of your task list will reason confidently about a world that is missing things. Ours is told it got a slice.

The tool descriptions tell the model when not to write. They are written to talk it out of speculative capture: file this if the person said to, not if it seemed like a good idea. The prose is part of the mechanism, not documentation added afterwards. It is the only instruction the model reliably reads.

Every feature we don't add is a decision to respect your attention. A connector is the largest opportunity to break that rule we have ever had.

Three moments

You are half an hour into a conversation about a migration, and somewhere in the middle you realise you need to talk to whoever owns the deploy pipeline before Thursday. You say so, in the same sentence you were already writing. The assistant captures it and the conversation carries on. The next morning it is on your new tab, in your own words, and you never opened Slaet to put it there.

You start the day by asking what is on. Your assistant calls whats_next and gets one task back. Not a ranked digest of nine, not a summary of your week. One task, or the fact that there isn't one. Both are useful answers, and the second one is the harder thing for software to be willing to say.

You keep a running note on a piece of work. Something worth recording comes out of a conversation with your assistant, so it searches your notes, finds that one, and appends a few lines to the end. Everything already in the note is exactly where you left it. You read the addition later, on a new tab, along with the rest.

On trust

The connector authenticates with OAuth. When it acts, it acts as you, carrying your own token. There is no privileged key sitting behind it with access to everyone's data.

Access is enforced by row-level security in the database, the same policies that govern the browser app. That distinction matters more than it sounds. If the connector's own logic were wrong tomorrow, if some tool handler forgot to check who was asking, the database would still refuse the write. The rules do not live in the part of the system we changed most recently.

Adding it

Open your assistant's custom connector settings, paste https://mcp.slaet.space/mcp, and sign in with your Slaet account. That is the setup. The first thing worth asking it is what you should be doing right now, and the best possible answer is that there's nothing.