Margin Intelligence

An assistant that operates the product through the same tools you do, what one action costs against your allowance, and where it is not allowed to look.

Checked against the running product on 7 min readMarkdown

On this page

Margin Intelligence is the assistant built into The Margin. Press ⌘J anywhere in the app to open it, or CtrlJ on Windows and Linux.

You can also talk to it. Tap the mic in the message box to dictate, or hold it to talk and let go: the words are sent after a short countdown you can undo. It works the same on the phone. See Talking instead of typing.

What it can do#

It reads and writes your workspace for you. A few of the things people ask it for:

  • Make a board, add and move cards, complete them, lay out a week.
  • Write, edit and search notes, or add to one you already have.
  • Log a habit, record an expense, some income or money someone owes.
  • Add shapes and text to a whiteboard.
  • Save a recipe from a link. In a family workspace, put dinner on Thursday, send the ingredients to the shopping list, and run chores and pocket money.
  • Read and reply in chat threads you are in.
  • Answer questions from what the Mind remembers.
  • Tell you what your week looks like, in a few lines. See Your week, across workspaces.
  • Explain how any part of The Margin works. It looks the answer up in these docs and gives you the link to the page, so "how do I add someone to the family?" gets the same answer the docs give.

It gets there with the tools an agent you connect over MCP gets, with a few of its own on top: searching the web, handing work to a specialist, and asking where to keep something you told it about yourself. It acts as you, with your role, so it cannot do anything you could not do yourself in the app. See Agents and MCP.

Your week, across workspaces#

Most people's week does not live in one workspace. The pitch is in your own, the school run and the grocery list are in the household's. Ask "what does my week look like?" and Margin Intelligence answers from all of them at once: tasks that are due, what is on the calendar, the dinners that are planned, chores and whose turn each one is, what is still on the shopping lists, your habits and your own journal lines. The answer is short, grouped by day: the workspace you asked from in full, and each other workspace as what is yours this week plus one line for the rest ("Studio North: 8 tasks overdue"). Ask about that workspace and it opens the line up. Two workspaces' shopping lists are reported separately, never as one count.

If you would rather keep them apart, open Settings→Your Brain on the web, or Settings→Data in the phone app, and set Answer from to This workspace. The choice follows your account to every device. It appears only if you belong to more than one workspace.

A few things hold whichever you choose:

  • Across workspaces it only reads. Anything it adds or changes still goes into the workspace you are in.
  • What it reads in another workspace is not remembered in this one. A conversation that looked across workspaces teaches the Mind nothing, so a housemate's pet or a team's client never turns up as a fact about you in your personal workspace.
  • Each workspace keeps its own rules. It reads boards you are a member of, leaves out locked notes and boards, your Vault and anything hidden from memory, and shows only your own habits and journal, never a housemate's.
  • Each workspace keeps its own plan. Chores, meals and the shopping list are read where that workspace's plan includes them.
  • The setting covers memory too. With All my workspaces, when it looks something up in memory it searches the Mind of each workspace you belong to, under that workspace's rules, and says where a fact came from ("from Home: the dentist is Dr. Lee"). With This workspace, only the one you are in.
  • A chore is reported as done only when it has been checked off. On a chore that takes turns it names whose turn it is now and who did it last time.

Assistants you connect yourself (Claude, ChatGPT, Cursor and others) follow this setting too, inside the workspaces you granted when you connected them. This workspace narrows them to the workspace they are working in; All my workspaces never reaches past what you granted. See Agents and MCP.

Wherever you ask, it also reads About you: the few facts about yourself that you keep for every workspace and nobody else sees. Tell it something about yourself in your own workspace and it keeps it there. In a household or team workspace it first asks Where should I keep this?, with For the household (or For the team) and Just for me, and nothing is saved until you choose.

Specialist agents#

The assistant hands a week's planning to a planner, and a question that needs digging to a research agent that searches the web. That works on every plan and needs nothing from you.

On the Team plan a team can also make agents of its own: a name, instructions and a set of tools, run by hand or on a schedule, with every result kept for the team to read. See Team agents. To work with an assistant you already use, connect it over MCP: see Agents and MCP.

When several agents work in one workspace, the rules they work under are a written pact. See Agent Pacts.

What an action costs#

AI actions a month

75

on Free

600

on Pro

1,200

on Family, pooled across the household

600

per seat on Team

100

while a trial runs, unless your paid plan already gives more

One action is a typical request. A turn that calls several tools is a loop of calls, so a longer piece of work counts as several actions. What you have used this month is in Settings→Billing.

When the allowance runs out, the assistant says so and stops. Everything else in the app keeps working. From there you can:

  • top up in Settings→Billing: $5 for 150 actions or $20 for 800, and top-ups never expire;
  • connect an MCP client of your own, which never draws on this allowance (Pro and up);
  • or wait for the count to start again on the 1st.

The Mind's background work has a separate allowance and never touches this one. See The Mind.

When it says it is paused#

The service has a daily limit on what the assistant can spend, everyone together. Free plans pause first, once the day's spending reaches their share of that limit (half of it), and everyone pauses only if the whole limit is reached. On Free your message then gets this answer in place of a reply: "Margin Intelligence is paused for free plans for the rest of today, because free plans have used their share of what the service can spend in a day. Paid plans keep working. It is back when the day resets at midnight UTC, or right away on a paid plan. Nothing you did is lost." If the whole limit is reached, everyone sees: "Margin Intelligence is paused for the rest of today, because the whole service hit its daily limit. It is back when the day resets at midnight UTC. Nothing you did is lost." The pause uses none of your own allowance. See Plans and limits.

What it is not allowed to see#

Your Vault while it is locked (and, through its tools, even while it is open), and anything marked Hide from memory. Both are kept out of retrieval itself, so they never reach an answer to be filtered out at the end. Something woven privately is recalled only when it is directly relevant, and never leads.

In a shared workspace the assistant sees what you can see. It cannot be used to read a household member's private material, or a board you are not on.

Where it runs#

The assistant runs as its own service, apart from the web app, so the rest of the product stays quick while a long task is in flight. It also means the assistant needs a connection. Everything else in the app works offline; this does not.

When it gets something wrong#

It will. If it keeps believing something untrue about you, find the line in Settings→Your Brain and edit or forget it. If it has a person or a project wrong, open them on Names in the Mind and teach it there. Telling it in the chat may or may not be remembered; a line you add to About you always is, in every workspace. An edit to the memory is read by every answer after it.