MCP Server

Groupthink's MCP (Model Context Protocol) server lets you ask any MCP-compatible AI assistant (Claude Desktop, Claude Code, Cursor, and others) to use Groupthink on your behalf. Your assistant can send the Groupthink bot into live meetings (joining, transcribing, speaking, chatting, leaving), read and write your relationship intelligence, and pull notes from meetings you have already had, surfaced in whichever tool you already work in.

New to this? Start at groupthink.com/agents for a two-minute overview and quickstart. This page is the complete reference: setup for every client, token management, and the full tool catalog.

• "Send the Groupthink bot to this Zoom and tell me what got covered when it's done: https://zoom.us/j/..."
• "What is happening in the meeting right now?"
• "Who am I meeting with this afternoon, and what should I prepare?"
• "Show me the notes from my call with sarah@example.com yesterday."
• "Show me my relationships I'm falling behind on"
• "What do I know about sarah@example.com?"
• "Save this as a note on sarah@example.com: we agreed to revisit pricing in Q3."
• "Find people whose record mentions pricing"

Prerequisites

Before setting up, you'll need:

• A Groupthink account, signed in at app.groupthink.com
• An MCP-compatible AI client. Tested with Claude Desktop, Claude Code, and Cursor; also works with other clients that support remote MCP servers over HTTP.

Two Ways to Connect

There are two ways to give your assistant access to Groupthink. Pick one.

The fastest way to set up either one is Connect an agent, at the top of MCP Access in Settings → Connections. Pick your agent from its tabs (Claude Code, Claude, ChatGPT, Cursor or Other apps) and copy what it shows: the finished command for Claude Code, the connector address for Claude and ChatGPT, or the Add to Cursor button and the Cursor config. For apps that need a token, Other apps has a Make a token button (see Option B).

	Sign in with Groupthink	API token
How it works	You paste one URL. Your browser opens, you approve, done.	You mint a token in Settings and paste it into your client's config.
Works with	Clients that support MCP connectors: Claude Code, claude.ai, Claude Desktop, Cursor.	Any MCP client, including older versions and clients without connector support.
You can choose read-only	Yes	No. A token always has full access.
Best for	Most people.	Scripts, automation, and clients that cannot do the sign-in flow.

Sign in with Groupthink is the recommended path. Nothing secret is copied or stored in a config file, and you can grant read-only access if you only want your assistant to look things up.

The server URL is the same either way:

https://api.groupthink.com/v1/mcp-server

Option A: Sign In With Groupthink

Claude Code

From your terminal, run:

claude mcp add --transport http --scope user groupthink https://api.groupthink.com/v1/mcp-server

No token, no header. The --scope user flag makes Groupthink available in every Claude Code project, not just the directory you ran the command from.

Then authenticate. Either run claude mcp login groupthink in your terminal, or run /mcp inside Claude Code, choose groupthink, and pick the authenticate option. Your browser opens on the Groupthink approval screen. Approve it, return to your terminal, and claude mcp list will show Groupthink as connected.

If you are having an assistant set this up for you, point it at claude mcp login: slash commands only run when a person types them.

claude.ai and Claude Desktop

Claude Desktop and claude.ai share the same connector settings, so adding it in one adds it in both.

1. Open Settings, then Connectors.
2. Choose Add custom connector.
3. Paste https://api.groupthink.com/v1/mcp-server as the URL and save. Leave the optional OAuth fields blank; Groupthink publishes what Claude needs to find on its own.
4. Click Connect. Your browser opens on the Groupthink approval screen.
5. Approve. The connector shows as connected and the Groupthink tools become available in your conversations.

You no longer need the mcp-remote bridge or a Node.js install for Claude Desktop. If you set Groupthink up that way previously, see Moving from a token to sign-in below.

Cursor

Open Cursor settings, go to the MCP Servers section, and add Groupthink as a remote HTTP server pointed at https://api.groupthink.com/v1/mcp-server with no authorization header. Cursor opens your browser for the same approval screen.

Other clients

Any MCP client that supports remote HTTP servers and OAuth can connect the same way: point it at https://api.groupthink.com/v1/mcp-server with no credential. The client discovers everything else it needs from the server. If your client asks you to supply an authorization URL or a client ID by hand, it does not support automatic discovery yet, and the API token path below is the one to use.

What You Are Approving

The approval screen names the app that is asking and shows what it will be able to do.

Groupthink cannot verify the name an app shows you. Any developer can register a connector and pick any name, including one that looks like ours. Only approve a connection you just started yourself, from the client you started it in.

You get a choice of what to grant:

• Read and write your Groupthink relationship data. The default, and what most assistants ask for. Everything in the tool catalog below, including saving notes and sending the meeting bot.
• Read your Groupthink relationship data. Read-only. The write tools are not restricted after the fact, they are invisible: an assistant holding a read-only grant cannot see them in its tool list at all, so it cannot call them by accident.

Read-only is a real downgrade you can pick at approval time. If you are not sure, start read-only. You can reconnect later with full access.

Option B: API Token

Use this for clients that cannot do the sign-in flow.

Mint a token

1. Open Settings → Connections on app.groupthink.com (sign in if prompted).
2. Find MCP Access, below the connected accounts and GitHub cards. The link in step 1 takes you there.
3. Click New token and give it a recognizable Token name (such as "Claude Desktop" or "Cursor on personal Mac"). If your account has a password, type it under Current password. Then click Create token. If your account has no password, Groupthink asks you to confirm it's you first, then creates the token.
4. Copy the token value when it appears, with Copy token (or Copy as Claude Desktop config for a ready-to-paste config). You will only see this once. Save it somewhere safe (a password manager works well). Once you click Done, the token cannot be retrieved.

A token always carries full read and write access. There is no read-only token; if you want read-only, use the sign-in path above.

Want the command filled in for you? In Connect an agent at the top of MCP Access, open Other apps and click Make a token. The token form opens with a name already filled in. Once you create the token, the tab shows the Claude Code command and the Claude Desktop config with your token in them, so you do not have to replace YOUR_TOKEN_HERE by hand. Copy them before you click Done.

Claude Code with a token

claude mcp add --transport http --scope user groupthink https://api.groupthink.com/v1/mcp-server \
  --header "Authorization: Bearer YOUR_TOKEN_HERE"

Claude Desktop and Cursor with a token

These reach Groupthink through a small bridge called mcp-remote, which runs automatically via npx on first use. You need Node.js installed (most modern Macs already have it; check by running node --version in Terminal).

Edit your config file:

• Claude Desktop, macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Claude Desktop, Windows: %APPDATA%\Claude\claude_desktop_config.json
• Cursor: the MCP Servers section of Cursor settings

Add the groupthink entry inside the existing mcpServers block. If you already have other MCP servers configured (for example, Playwright or a Cooking MCP), keep them. Add the groupthink key alongside them, don't replace the whole file:

{
  "mcpServers": {
    "groupthink": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://api.groupthink.com/v1/mcp-server",
        "--header",
        "Authorization:Bearer YOUR_TOKEN_HERE"
      ]
    }
  }
}

Replace YOUR_TOKEN_HERE with your token. Save the file and fully quit and relaunch the app.

Moving from a token to sign-in

If you already connected with a token and want to switch:

1. Remove the existing groupthink entry from your client (for Claude Code, claude mcp remove groupthink; for Claude Desktop or Cursor, delete the groupthink block from the config file).
2. Follow Option A for your client.
3. Revoke the old token under Settings → Connections → MCP Access so it stops working.

Both methods can coexist while you switch, so nothing breaks in between.

Try It Out

Note: Following a live call from your assistant isn't working right now. The bot can be sent, but your assistant can't read the call while it's going, and no transcript is saved afterward. Reading recaps and transcripts after a meeting still works for meetings Groupthink joined from your calendar or that you recorded with the desktop app. Ask something like "What came out of my call with Theo?"

Open a new conversation in your AI assistant and try a natural-language prompt. Sending the bot into a live meeting (have a Zoom, Google Meet, or Microsoft Teams URL ready):

• "Send the Groupthink bot to this Zoom: [your meeting URL]"
• "What is being said in the meeting right now?"
• "Have the bot say 'Sorry I'm running late, I'll be there in five minutes' in the meeting"
• "Tell the bot to leave the meeting"
• "Save these notes from the meeting: ..."

Working with your relationship intelligence:

• "What relationships do I have in Groupthink?"
• "Tell me about my relationship with [their email]"
• "Who am I falling behind on?"
• "What meetings do I have this week, and what should I know about the attendees?"
• "Show me the notes from my call with [their email] yesterday."
• "Save this as a note on [their email]: [paragraph of context, background, political read, etc.]"
• "Remind me to send [their email] the deck by Friday."

Your assistant should call the appropriate Groupthink tools and respond with your real data. Tools that change state (sending the bot, posting chat, speaking, saving notes) write a usage row to your account so you can see what was done on your behalf.

When Your Assistant Uses Groupthink on Its Own

You do not have to mention Groupthink for your assistant to use it. When your assistant connects, Groupthink gives it standing instructions to use Groupthink "whenever a person comes up, not only when the user mentions Groupthink." So your assistant may call Groupthink tools on its own when a request is about a person. That includes a person named in a transcript, email or notes you share with it.

If your agent does not pick up these instructions on its own, copy them from Standing instructions, below the tabs in Connect an agent, with its Copy button. Paste them into the agent's instructions file, such as AGENTS.md or CLAUDE.md.

What you may notice:

• Lookups. The assistant is told to look the person up with search_relationships and get_relationship before it answers, and to say so when nobody matches.
• Saves. On a read and write connection, the assistant is told to save new things it learns about a person it found in Groupthink, without being asked. It is told to then say in one short line what it saved and for whom. The tools it uses for this are stage_open_thread, capture, log_interaction and stage_dossier_entry.
• New people. When the lookup does not find someone, the assistant is told to search again under another name, and to ask you before saving anything about them, because saving would add them as a new contact.

These are instructions, and your assistant decides when to follow them. Check what it tells you it saved.

To stop the saves, connect read-only (see What You Are Approving). A read-only connection gets the lookup instructions only, and the assistant is told it can read your record but not change it. To stop the lookups as well, disconnect.

Reference: Tools Your Assistant Can Call

The MCP server exposes 35 tools across three areas: reading your relationship record, writing back to it, and participating live in meetings (the live-meeting tools are available when live meeting tools are enabled). You usually don't need to know these by name — your assistant picks the right one from a natural-language prompt — but here's the full catalog.

Live meeting participation

These tools send the Groupthink bot into a live meeting and let your assistant interact with it. Each tool except join_meeting operates on a bot_id returned by join_meeting.

Note: Following a live call from your assistant isn't working right now. The bot can be sent, but your assistant can't read the call while it's going, and no transcript is saved afterward. Reading recaps and transcripts after a meeting still works for meetings Groupthink joined from your calendar or that you recorded with the desktop app.

For a step-by-step walkthrough of following a call and asking about it as it happens, see Follow a Live Call With Your AI Assistant.

join_meeting, speak, send_chat, leave_meeting and save_notes need a read and write connection. get_transcript, bot_status and get_notes also work on a read-only one.

Tool	What it does
join_meeting	Send the Groupthink bot to a Zoom, Google Meet, or Microsoft Teams URL. Returns the bot_id that subsequent tools operate on.
get_transcript	Fetch new transcript lines from the bot since the last call. Maintains a per-user cursor so repeated calls return only what is new. For a past recording listed by get_past_meetings, returns the stored transcript.
speak	Have the bot speak text aloud in the meeting using TTS. Max 1000 characters per call. Subject to a small per-user daily spend cap to prevent runaway audio cost.
send_chat	Send a chat message into the meeting via the bot. Max 2000 characters.
bot_status	Get the bot's current state (joining, in_call_recording, done, errored, etc), name, meeting URL, and creation time.
leave_meeting	Tell the bot to leave the meeting. Recall may take a few seconds to remove the bot. Also clears the per-user transcript cursor.
save_notes	Persist agent-generated meeting notes (summary, action items, key decisions, contributions) against the bot's meeting. Idempotent. You can read them on the meeting's recap page in Groupthink.
get_notes	Retrieve the notes your own agents saved for the bot's meeting. Notes are private to you: other attendees can't read them.

Notes your agents save also appear in Groupthink, on the meeting's recap page. Open the meeting and look on the Summary tab, after the recap: each agent gets its own block labelled "Notes saved by" and the agent's name, with its summary, action items, key decisions and contributions. They're kept apart from Groupthink's own recap so you can always tell who wrote what, and only you can see them. On the iPhone app, the same notes show on the meeting's screen, below the recap.

What join_meeting accepts. When live meeting tools are enabled, join_meeting takes two parameters:

• meeting_url (required): the full join link for a Zoom, Google Meet or Microsoft Teams meeting. A meeting ID or a meeting name is not enough. join_meeting refuses it, sends no bot and creates nothing.
• bot_name (optional): the name the bot shows in the meeting. When your assistant leaves it out, the bot uses the Display name you set under Settings, Meetings, Bot identity. If that is blank, the bot joins as your name followed by "'s notetaker", for example Dana Lee's notetaker.

So you can ask for a name in the request, for example "Send the Groupthink bot to this Zoom as Dana's notetaker". join_meeting has no parameter for the bot's picture.

When join_meeting is refused. Sending the bot to a new meeting needs an active trial or subscription. Without one, join_meeting sends no bot and your assistant gets this message back, followed by an upgrade link:

No bot was sent and nothing was created. Recording a new meeting needs an active Groupthink trial or subscription, and this account does not have one right now. Past meetings and their recaps are still available.

The upgrade link takes you straight to checkout for Groupthink Pro, with no sign-in or plan picking first. It works for an hour. If you upgrade within 30 minutes, the bot is sent to that meeting automatically. If you already pay, or your card needs updating, the link opens your billing settings instead.

Your assistant can check ahead of time with get_capture_settings, which reports whether your plan allows new captures.

Your relationship and meeting intelligence

Read tools query the graph; upsert_relationship_note writes a free-form note back to it. get_upcoming_meetings and get_past_meetings cover the calendar in both directions so your assistant can answer "who am I meeting with at 2pm" and "what got covered in my call with sarah yesterday" without you remembering any IDs. Personal details surfaced by get_relationship carry source-meeting attribution + extraction timestamps so your assistant can verify currency before relaying a fact. Contact facts (mailing address, salutation, partner, anniversary) come with where they came from, when they last changed, and whether you've confirmed them, so your assistant can tell an address you checked from one an import brought in.

complete_action_item and upsert_relationship_note need a read and write connection. The other tools in this table also work on a read-only one.

Tool	What it does
list_relationships	Lists your relationships with frequency status, and which contact facts (mailing address, salutation, partner, anniversary) you have on file for each. Optional filter by status (on track, falling behind, overdue).
get_relationship	Returns detailed information about one relationship, looked up by email or relationship ID. Includes contact facts (mailing address, salutation, partner, anniversary, each with its source, when it last changed, and whether you confirmed it), personal details (with source meeting + extracted-at timestamp), saved facts, open threads, recent meetings, tags, user notes, and meeting context. Supports filter / pagination on personal_details.
search_relationships	Searches across names, emails, tags, and personal details to find specific contacts. Each result says which contact facts you have on file for that person.
get_relationship_insights	Returns a high-level summary of your relationship health and who needs outreach.
get_account_status	Reports the state of your Groupthink account: who the connection is signed in as, your plan and trial status, whether the notetaker may join new meetings, which calendars are connected, your note-taking preferences, whether this connection is read-only or read and write, and what is still missing from setup.
get_upcoming_meetings	Lists upcoming meetings with the relationship context for each attendee, including each attendee's relationship ID so your assistant can chain directly to get_relationship. Each meeting also says whether the Groupthink notetaker will join it (will_join, will_not_join or dismissed), the rule that decided it (for example organizer_only, private_recording, no_video_link, or meeting_filter for a task block, a block only you are on that you made), and why, so your assistant can tell you before the call if nothing will be recorded.
get_capture_settings	Reads the settings that decide whether your meetings are recorded: whether your plan allows new captures, the notetaker's display name, and for each connected calendar the auto-join mode and the filters behind it. Read-only.
get_past_meetings	Lists meetings you attended in the last N days (default 7, max 30, optional email filter). Every meeting includes its meeting_id, plus the Groupthink bot_id when a recording exists, so your assistant can chain into get_meeting_recap, get_notes or get_transcript without you remembering an id.
get_meeting_recap	Reads the recap for a past meeting: summary, narrative, decisions and intro brief, plus whether the recap has finished generating. If you captured a window during the meeting with the Mac app, it also includes what was shown on screen, or says why the capture found nothing. Only your own capture, never another attendee's. Takes the meeting_id from get_past_meetings. Works whether or not the Groupthink bot recorded the meeting. Returns shared (the group recap) unless your assistant asks for personal (your own version), so the version written just for you reaches a connected assistant only when it asks for it. Also includes any notes your own agents saved on the meeting with save_notes, each labelled with the agent that wrote it and kept separate from the recap.
list_action_items	Lists the action items Groupthink captured for you from your meetings and notes. Defaults to open items, soonest due first, in your timezone. Each item shows its ID, due date, the person it involves, and where it came from.
get_action_item	Returns the full context behind one action item: text, due date and status, the person it involves with their email, the quote it came from, and the meeting recap or note it was captured in.
complete_action_item	Marks an action item completed, the same as checking it off in Groupthink. For a task handed to your assistant, it can also send a short result and links (see Project tasks). Safe to call twice. It does not record whether you told the other person; the response says when that is still outstanding.
list_gift_ideas	Lists the gift ideas saved for a person: the idea, where it came up, and the occasion it is for. Ideas not yet given come first. Ones already given follow with the date and, when you recorded it, how the gift landed.
upsert_relationship_note	Writes a free-form, user-authored note onto a relationship. Stored verbatim, distinct from auto-extracted personal_details. Supports upsert by source so re-pushing the same content doesn't duplicate.

Meeting prep and plans

Your assistant can brief you before a meeting ("brief me for my 2pm") and keep track of the plans you are making with people ("what plans do I have with Dana?"). It sees what the app shows you.

propose_plan needs a read and write connection. The other three also work on a read-only one.

Tool	What it does
get_meeting_brief	The prep brief for one meeting. Your assistant passes the meeting ID from get_upcoming_meetings, or a person, for your next meeting with them in the coming 30 days. For each person in the meeting: what to raise (open threads), the goals in play, whether the relationship is drifting or overdue, when you last talked, and the most recent things you know about them.
list_plans	Lists your plans: who each one is with and where it stands (Proposed, or Time picked), what it is for, the dates it could fall on, and the time you picked. It can list only the plans with one person, or your archived plans.
get_plan	Returns one plan: everything above, plus when the other person said they are free (in their words, when Groupthink has them) and, until you pick a time, times that work for both of you, checked against your calendars. If a calendar needs reconnecting, it says that calendar was left out.
propose_plan	Starts a plan with one of your contacts, with the person in Proposed. Nothing is sent to them and nothing goes on your calendar. You pick the time in Groupthink, which is what puts it on your calendar and drafts a reply for you to send. If you already have an active plan with that person under the same name, it is not made twice.

A plan your assistant starts is marked as made by a connected assistant, with the connected app's name.

Writing back to your record

Beyond the free-form upsert_relationship_note, your assistant can maintain the structured parts of a relationship: the saved facts (durable facts and context) and open threads (the loops and commitments still in play). Every write authenticates as you and passes the same permission checks as the app, so an agent can only touch your own record.

Your assistant can also file a quick capture, log a conversation that already happened, save a gift idea, and add or edit action items. Every tool in this table needs a read and write connection.

Tool	What it does
stage_dossier_entry	Adds a new saved fact (a fact or piece of context) to a relationship.
refine_dossier_entry	Edits the text of an existing saved fact.
merge_dossier_entry	Merges duplicate saved facts into one.
archive_dossier_entry	Archives a saved fact so it no longer surfaces.
dismiss_dossier_entry	Dismisses a suggested or auto-extracted fact you do not want kept.
stage_open_thread	Opens a thread on a relationship (an open loop, commitment, or follow-up).
transition_open_thread	Moves a thread to a new state: covered, deferred, or no longer relevant. It can also reopen a closed thread.
update_open_thread	Edits the topic of an open thread, or the outcome recorded on a covered one, without changing its state.
ingest_meeting_content	Ingests pasted notes or transcript content as a capture, the same path the app uses for notes.
capture	Files one line in your own words (a fact about someone, a note to self, or a reminder) onto the right record, the same way the capture box in the app does. Up to 4,000 characters. The result says where it was filed, whether that is a guess you should confirm, and whether it added anyone as a new contact because no existing contact matched.
log_interaction	Logs a conversation that already happened with one person (a call, text, email, in-person meeting or gift) so it counts as contact and resets how long it has been since you were in touch. It can be backdated but not set in the future. With no channel given, it is saved as a private note and does not count as contact.
add_gift_idea	Saves a gift idea for a person, with where the idea came up and, if you named one, the occasion.
create_action_item	Adds an action item on a person, on a meeting you were on, or both: "remind me to send Sarah the deck". It shows on that person's page and in list_action_items. With a person, it counts as something you owe them unless your assistant says they owe you. It can set a due date and file it the way the app does (including opening a GitHub or Linear issue, or handing it to your assistant, which shows as Your assistant in the app). It needs a person or a meeting.
update_action_item	Changes an action item's wording, due date or where it is filed, the same edit as in the app. Filing it to your assistant keeps it open on your list until it is marked done. To check one off, your assistant uses complete_action_item.
attach_action_item	Puts an existing action item on a person (it then shows on their page) or on a meeting you were on (everyone who can see that meeting sees it there). An item already on one meeting is not moved to another. Safe to call twice.
merge_relationships	Merges two contacts that are the same person, after you say yes. See below.

Groupthink records what your assistant wrote

When your assistant opens a thread with stage_open_thread, logs a conversation with log_interaction, adds an action item with create_action_item, or saves or rewords a fact about someone with stage_dossier_entry or refine_dossier_entry, Groupthink records that a connected assistant wrote it, and the name of the connected app (for example "Claude Code", or the name you gave an API token). Captures from capture already carry the app's name. On the web, the person's page says so for threads, logged conversations and saved facts: each reads Added by and the connected app's name exactly as the app gave it when it connected, for example Added by Claude Code (groupthink), or Added by an agent when no name was recorded. See Relationship Intelligence for threads and Details for facts. It is also in your account data export. The iPhone app does not show it yet.

The app's name is the one the app gave itself when it connected. Groupthink does not verify it.

Threads and conversations your assistant added before October 12, 2026 are not marked, because nothing was recorded about who wrote them then. The same goes for facts saved before this change.

Merging two contacts that are the same person

If your assistant notices two records for one person (say "Kate" and "Kate Morrow"), it can offer to merge them. Nothing changes until you agree. It first shows you both names, which one is kept, and what would move over: conversations, meetings, facts, open threads, emails and phone numbers. Only after you say yes does it merge them. Your assistant can't undo a merge, and it may tell you a merge can't be undone. You can undo it in the app: open the contact that was kept and click Undo merge, which works until either contact is edited. See Undo a merge. A read-only connection can't merge; it will point you to the app instead.

Your projects

Projects are how you group people and meetings around one effort: a deal, a client, a trip, a holiday card list. Your assistant can read them and keep them up to date, so you can ask "what's left on the Acme project?" or say "add Sarah to the Acme project". It can name a project by its name, in any capitalization, or by its ID.

list_projects and get_project also work on a read-only connection. The other four need a read and write connection.

Tool	What it does
list_projects	Lists your active projects with their kind (work, personal, family or plan) and purpose. It can list your archived projects instead.
get_project	Returns one project: its purpose, the people on it (people added to it, with who added each one: you, an assistant by the name it gives, or a plan you made; and people whose notes or meetings are filed on it), how many meetings are filed on it, and its open action items, soonest due first. Muted people are left out.
create_project	Makes a new project. If you already have an active project with that name, it uses that one instead of making a second.
update_project	Renames a project, or sets what it is for. A purpose your assistant sets counts as one you confirmed.
add_person_to_project	Puts a person on a project. For someone already in Groupthink, your assistant passes their email or ID. For someone new, it passes their name (and email, if known) and Groupthink adds them as a contact. Groupthink notes that your assistant added them, and which app it was. Adding someone twice changes nothing, and someone you added stays marked as added by you.
remove_person_from_project	Takes a person off a project. Nothing about the person is deleted. If notes or meetings about them are filed on the project, they still show on it.

A project your assistant creates is marked as made by a connected assistant, with the connected app's name, the same way its threads and logged conversations are. So is a contact it adds by name.

Project tasks

To read a project's task list, your assistant calls list_action_items with project set to the project's name or ID, the same names list_projects gives. It lists only the tasks on that project, open ones by default. On a task of your own that you are not doing yourself, the line says who does it (for example doer: agent (claude)) and, once work on it has started, how it is going (for example run_status: waiting_on_agent).

A task can be handed to ChatGPT, Claude or Dots to do for you. Claude and Dots pick their tasks up from Groupthink over this connection:

1. The assistant calls list_action_items with handed_to set to chatgpt, claude or dots. It lists only the tasks handed to that assistant.
2. get_action_item on one of them shows which assistant it was handed to and how the run stands, and tells the assistant to report back when it is done.
3. When it has done the task, the assistant calls complete_action_item with result, a sentence or two on what it did, and links, up to 10 http or https links to what it made (a draft, a document, a booking). The task is checked off, and a task that was waiting on the assistant is marked done, with or without a result.

A result sent for a task you already checked off still counts once; sending it again changes nothing.

Disconnecting

A connection you made by signing in: remove the connector in the client that holds it. In Claude Code, claude mcp remove groupthink. In claude.ai or Claude Desktop, open Settings, Connectors, and delete the Groupthink connector. In Cursor, delete the entry from the MCP Servers list.

You can also disconnect it from Groupthink, which works even if you cannot reach the client (a shared machine you no longer have, or a connection you do not recognize). Open Settings, switch to the Connections tab, and find the app under Connected agents in MCP Access. Click Disconnect, then Disconnect again to confirm. Access ends immediately when it is cut, not at the end of the grant.

On iPhone, open the Settings tab and tap AI agents using your account, under Connected agents. Tap Disconnect next to the agent, then Disconnect again to confirm. The row only shows while you are signed in to Groupthink. See Connected agents.

A connection you made with a token: open Settings, switch to the Connections tab, find the token under Your tokens in MCP Access, and click Revoke. Click Revoke again to confirm. Any client using the revoked token will immediately lose access. Revoke any token you no longer recognize, or that you saved somewhere insecure.

How Long a Connection Lasts

A sign-in connection refreshes itself in the background and lasts up to six months of continued use before your client asks you to sign in again. Tokens are separate: a token you minted keeps working until you revoke it.

Troubleshooting

"401 Unauthorized" or "Authentication required" If you connected with a token, it is missing, expired, or revoked. Mint a new one and update your client config. If you connected by signing in, the connection was disconnected or has lapsed; remove it from your client and add it again.

Your client asks you to paste an authorization URL, token URL, or client ID It does not support automatic discovery. Leave those fields blank if it lets you, or use the API token path instead.

Your client says the connection was refused when you approved it Some clients ask you to sign in from a browser profile that is not signed in to Groupthink. Sign in at app.groupthink.com in the same browser, then try connecting again.

Your assistant cannot see the tools for saving notes or sending the bot The connection was approved as read-only. Remove it in your client, add it again, and pick read and write on the approval screen.

"No relationships found" or empty responses You may not have relationships tracked yet. Relationships are built automatically from your meeting activity in Groupthink. Run a meeting or two first.

"No calendars connected" The get_upcoming_meetings tool needs a connected calendar. Connect Google Calendar or Outlook in Groupthink first.

Claude Desktop says no MCP servers found, or the Groupthink tools don't show up If you connected by signing in, check Settings, Connectors, and confirm Groupthink shows as connected. If you used a token, make sure your claude_desktop_config.json is valid JSON and that you fully quit and relaunched Claude Desktop after editing it.

What's Next

• Manage your tokens under Settings → Connections → MCP Access (direct link). You can mint and revoke tokens from there at any time.
• More tools will land as Groupthink's relationship intelligence grows. If a tool is missing that would unlock a workflow for you, email support@groupthink.com.

Privacy & Scope

However you connect, your assistant authenticates as your user account and reaches the same data you do in the Groupthink app, scoped to your relationships and your calendars. Neither method grants access to other users' data.

A sign-in connection can be narrowed to read-only at the approval screen. A token cannot; it always carries full read and write access, so treat your tokens like passwords. Anyone with a token can read your relationship data through any MCP client.

Tools that change state (sending the bot, posting chat, speaking, saving notes) write a usage row to your account so you can see what was done on your behalf.