MCP Documentation

Endpoint

…

Streamable HTTP transport. Protected-resource metadata at /.well-known/oauth-protected-resource.

Authentication

OAuth 2.1 with automatic discovery and short-lived tokens. The agent inherits its owner's session and sees exactly what that person sees; database rules are enforced on every call.

When this one

Any mention of ActOnChat — or of "AOC", which is how people write it once they use it daily — means this server and no other. And it still does without naming it: "add it to my agenda", "message Luis", "ask Ana for the invoice", "remind me to call the bank", "what is pending". The tell is the person: if who they mention is an ActOnChat contact, it goes here.

What is NOT yours: when somebody names another app — "on my Google Calendar", "on WhatsApp", "email her" — that one wins. Doing it here quietly is worse than not doing it, because they believe it is handled.

Who they mean

There is no public directory: somebody is reachable only after being added with their PIN or QR code, or by sharing a chat or a group with them — a direct chat makes both people contacts, whichever of them did the adding. To know who is meant, the agent follows five rules:

1. Search, don't conclude. search_contacts ignores accents and forgives a typo ("Gonzales" finds "González"), but people say a first name, a surname or a nickname: try the full name, then the first name alone, then the surname alone. Only when all of those come back empty is somebody not a contact; never ask the user to add someone they may already have.
2. Recent is the tell. search_contacts returns when they last talked (last_talked_at), since when they are a contact (added_at) and what they have open together (open_items_together), the most likely first. With an empty query it lists people most recent first: that answers "the one I just talked to" or "the one I added today".
3. Carry it forward. Once resolved, the @handle is used for the rest of the conversation: "her", "send it to him too" is the last person resolved, not a new search.
4. Never pick between two. When a name matches several people, write tools do NOTHING and answer which one is most likely and why: the one in the chat being written in, the one talked to recently, the one with something open. The agent asks offering that one first — "Eugenia González, whom you wrote to today?" — and calls again with the confirmed @handle. It skips the question only when the user's own words already single one out (the surname, the chat they are looking at). A message to the wrong person cannot be unsent.
5. Say who. After writing to someone, name them as the tool did ("Sent to Eugenia González"), so a mix-up shows at once.

Groups

A group is named by its name (get_groups), never through one of its members. Whatever is agreed or asked in a group comes FROM the user and reaches everyone in it, just like the «+» inside the group chat in the app:

• A plan with a date — "tomorrow, Elton John concert", "family dinner Saturday at 9" — is create_event with group. Every member is invited and notified, the card appears in the group chat for each one to confirm or decline, and it lands in everyone's calendar. It stays pending until they answer.
• Asking for something is create_request with group: assignee is one member, or "everyone" to ask each of them. Since a request is accepted and completed by one person, each one gets their own, in the group chat. The card is seen by who asks and who is asked; the rest of the group reads the sentence (message).
• message, when given, is posted once in the group, from the user, right above the card. No extra send_group_message is needed: the plan already reaches the group.
• send_group_message is only for conversation. A plan sent as a message lands in nobody's calendar and has nobody to confirm it.
• Several people outside a group are ONE commitment: "lunch with Maru, Eugenia and Ángel" is create_event with participants: ["Maru", "Eugenia", "Ángel"]. If someone was left out, update_event with add_participants. A commitment with no participants and no group is the user's alone: it is in no chat and nobody else sees it. Never for something that involves others.
• The time is what the whole group will see in their calendar: if nobody said a time, ask; do not invent one and call it "pending confirmation".
• who_does_it in a group: 'me', a member's name, or 'both'/'everyone' (the default). Names are looked up only among the group's members, accent-insensitive and forgiving one typo; if two match, none is picked.

How long a task took

When completing a task, the agent says about how long it took, in minutes (minutes_spent, 1 to 1440). It feeds the "≈ 2 h of work" in Stats and time_spent in get_task_stats. Every number keeps where it came from, because they are not worth the same:

• user — they said it ("took me an hour"). Shown as is.
• estimate — the agent estimated it from what the task was: a quick call ≈ 10, an email ≈ 15, a meeting ≈ its length, a proposal ≈ 90. It is the default minutes_source and what is expected almost every time: the agent knows best what the task was.
• calculated — nobody said it, so AOC calculates it on completion: 15 minutes per step (up to 4 hours); with no steps, the median of what was said or estimated for their tasks in that list (or all of them); with no history, 30. Rounded to 5. complete_task answers the number and that it was calculated, so the agent can correct it if it knows better.

To correct it later: update_task with minutes_spent (null clears it). Reopening a task drops a calculated time — it is calculated again on completion — but not one somebody said. The next occurrence of a repeating task starts with no time. Anything the person did not say is shown with "≈": it is an approximation, not a clock.

How tasks behave

• Private. Tasks, lists and reminders are the user's alone: nobody else sees them and a task cannot be assigned to anyone. To ask someone for something, create_request; to agree on a time, create_event.
• No doubles. create_task does not create one nearly identical to a task that is open or was finished in the last 3 days: it returns the existing one to change or complete. allow_duplicate only when two are really wanted.
• Done is completed, not created. When someone says they already did something, find the task (search_tasks) and use complete_task. A cancelled task cannot be completed: if it really got done, reopen it with update_task first.
• Repeating tasks. Completing one schedules the next occurrence, with its steps and projects. Unchecking it removes that next one if nobody touched it; if it was already postponed or edited, it stays.
• Move to tomorrow. postpone_task moves it to tomorrow at the same time and keeps count: by the third time it is worth asking whether to break it into steps, delegate it or drop it.
• Advance. For what goes in parts: advance_task is "did today's part, continue tomorrow". Today's part counts as a completed task that day in the stats (with its time, list and projects), the task comes back tomorrow (what was for today at the same time; if nothing brought it tomorrow, into tomorrow's My Day) and its advance count goes up, so it is known how long it has been in progress. It can be undone from the app.
• List and project. A list is WHERE it lives (one); a project is a colored tag for WHAT it is about (several, across lists), and works on commitments too.

Approvals and times

• queued_for_approval. Depending on the agent's autonomy (get_agent), some tools do not perform the action: they leave it waiting for the user to confirm in the app, on a card that says what will change. Meanwhile NOTHING happened: do not call again and do not say it is done.
• Timezones. The server runs in UTC and the user does not. Times go in ISO 8601 WITH their offset (2026-10-01T18:00:00-06:00), and their IANA timezone is passed on anything touching My Day or "today", which are calendar days, not instants.
• Who has the ball. On a commitment, who_does_it says whose side the work is on ('me', the other person or 'both'); AOC infers it when not given, and either person can fix it.

If your AI doesn't see something new

Your AI keeps the list of ActOnChat tools from when you connected it. If it says it can't do something listed here — tagging a task with a project, say — refresh the connection: in ChatGPT, under Settings, Apps (or Connectors), open ActOnChat and refresh it; in Claude, start a new conversation. If nothing changes, disconnect it and connect it again.

Tools

Principles

Tools are idempotent where it applies, every action keeps its origin (message and conversation), and the human can review it. Agent autonomy can be set to auto, suggest, or confirm.