Loopr Docs App →

MCP Server

What this is for

Loopr ships a built-in Model Context Protocol server. Point an MCP-aware AI client at it and the assistant can read your loop, complete or snooze items, check off your morning routine, log an activity, jot a note on a person you track — anything you’d otherwise do in the UI, scoped to your account.

It’s the same idea as the GraphQL API, but the AI handles tool discovery, auth, and request shape on its own. You sign in once; from then on the assistant just calls tools.

Typical uses:

  • “Snooze everything from GitHub until tomorrow morning.”
  • “Complete the things I’ve already done this morning and tell me what’s left.”
  • “Add a talking point for my 1:1 with Sam: ask about the migration.”
  • “Log that I went for a run.”

Endpoint

https://loopr.life/mcp

Auth is OAuth 2.1 with dynamic client registration — most MCP clients will just ask for the URL and walk you through the consent screen the first time you connect. The token is scoped to your account and acts as you, the same way an API key does.

If your client needs them explicitly, the discovery endpoints are:

  • https://loopr.life/.well-known/oauth-authorization-server
  • https://loopr.life/.well-known/oauth-protected-resource

The required scope is mcp.

Connecting from Claude Desktop

In Claude’s settings, add a new Connector and paste https://loopr.life/mcp as the URL. Claude will pop a browser window for you to sign in to Loopr and approve the connection. After that, all the tools below are available in any chat.

(The exact menu name varies a bit between Claude clients — look for “connectors”, “MCP servers”, or “remote tools”.)

What the AI can do

Tools are grouped roughly the way the app is.

The loop

  • list_queued_items, list_snoozed_items, list_holding_items, list_top_of_mind_items, list_urgent_items — read what’s in there.
  • list_completed_items — paginated history of completed items, newest first.
  • search_items — substring search across all items in any state, with an optional state filter. Required pagination keeps the AI from pulling your entire archive at once.
  • create_item — push something new into the loop.
  • complete_item, snooze_item, unsnooze_item, requeue_item — move items between states.
  • set_item_urgency, toggle_top_of_mind, promote_from_holding — adjust priority and pinning.
  • list_streams — see your wired-up integrations.

Lookbacks

  • browse_lookbacks — paginated browse of your retrospectives, scoped to a single period. Requires a period argument (daily, weekly, monthly, or quarterly).
  • get_lookback — fetch a single lookback by id (e.g. to read its full body after browsing).

Daily routine & activities

  • list_routine_items, list_todays_routine_items — see your morning / evening routine.
  • create_routine_item, update_routine_item, delete_routine_item — manage what’s on it.
  • check_routine_item, uncheck_routine_item, list_todays_checks — tick today’s items off (or undo).
  • list_activities, create_activity, update_activity, delete_activity — manage the menu of activities you log.
  • log_activity, unlog_activity, list_todays_activity_logs — record a completion (“I went for a run”).

People

  • list_people, create_person, update_person, destroy_person — the people you track.
  • update_person_public_notes, update_person_private_notes — keep notes (public ones show on the share link; private ones don’t).
  • list_open_delegations, list_delegations_for_person, create_delegation, update_delegation, update_delegation_status, destroy_delegation — what you’ve handed off to whom.
  • list_agenda_items_for_person, create_agenda_item, complete_agenda_item, uncomplete_agenda_item, destroy_agenda_item — talking points for your next 1:1.

The schema for every tool is exposed by the server itself, so you don’t need to memorize argument names — the assistant fetches them on demand.

Revoking access

Tokens issued to MCP clients live in Preferences → Connected apps. Revoke any client there and its tokens stop working immediately. The client will need to re-do the consent flow to reconnect.

Non-goals

  • No fine-grained scopes yet. A connected client can do anything you can. Treat the consent screen the way you’d treat handing someone your laptop.
  • No webhooks out of Loopr. The MCP server is request/response only; the AI calls tools when it needs to, it doesn’t subscribe to events.

Compared to the API key flow

The GraphQL API is still the right tool for scripts, Zapier, and anything programmatic — long-lived API key, one URL, one schema. MCP is the right tool for an AI assistant: short-lived OAuth tokens, structured tool descriptions the model can introspect, no hand-rolled prompts about “here’s how to call the API”.

Back to app