connect Claude, ChatGPT, Cursor and friends

MCP

init.Tasks speaks MCP at https://mcp.inittasks.com/mcp, so an AI client can work with a user's tasks directly — the same data, the same scopes, the same revocation.

It is the same server and the same encryption as the REST API. An MCP client is just another OAuth app.

Connect

Most clients take a URL. In Claude Code:

bash

claude mcp add --transport http init-tasks https://mcp.inittasks.com/mcp

In a client that reads a JSON config:

json

{
  "mcpServers": {
    "init-tasks": {
      "type": "http",
      "url": "https://mcp.inittasks.com/mcp"
    }
  }
}

Then the client opens the approval flow. You approve on your phone, not in the client's browser window: init.Tasks shows a nine-digit code, you open the app on a device you already trust, and confirm there. No password is typed into anything but the app itself.

Read-only by default

A connection asks for read access unless you grant more. 12 of the 39 tools work read-only; the rest need write access, and the server refuses them without it — the refusal is enforced server-side, not left to the client to respect.

Revoke any time in Settings → account → encryption & devices → connected apps. A password change or an encryption-key change disconnects everything at once.

The tools

tool	read-only	what it does
`archive_inbox_item`	[ ]	Archive an inbox item (hides it from the inbox).
`capture_inbox`	[ ]	Quick-add capture into the inbox (v1.2). Kind (text/url) is auto-detected. For a text capture the trailing date phrase is parsed via the canonical quick-add grammar ("call maria tomorrow", "renew passport 7 oct") — the remainder becomes the item text and the parsed date is PERSISTED on the item, ready to ride onto the to-do when it is later moved into a project. A url fires the title fetcher best-effort.
`complete_todo`	[ ]	Mark a to-do done (idempotent). A recurring active occurrence runs the spec/60 transition — a done copy is created and the active row advances to its next occurrence. Resolve by id or exact title.
`create_area`	[ ]	Create a top-level area.
`create_filter`	[ ]	Create a saved filter (v1.4). `criteria` is { v:1, all:[ …criterion ] }, AND-composed. A criterion is one of: {kind:"tag",name} · {kind:"priority",min:"low"\	"medium"\	"high"} (matches priority ≥ min) · {kind:"deadline",withinDays:0..365} (deadline ≤ today+N, overdue included) · {kind:"deadline",overdue:true} · {kind:"scheduled",state:"dated"\	"anytime"\	"someday"} · {kind:"where",containerId} (the to-do’s ancestor chain contains that area/project/subproject id). An unknown/newer-version criterion is rejected here (only stored filters may be inert). Returns the filter and its match count.
`create_note`	[ ]	v1.6: create a standalone NOTE in the inbox — content the user means to KEEP, as opposed to a scrap awaiting triage. Stored as tasks_inbox.kind='note' with the body in `text` and an optional heading in `title` (spec/10 "Notes"), so it lives alongside attachments and inherits the whole inbox pipeline. ⛔ A note is NOT a to-do: it has no done state, never appears in a date view, and counts only toward `inbox`. Kind is never auto-detected for notes — use capture_inbox for ordinary text/url captures.
`create_project`	[ ]	Create a project. area is resolved by name or id; omit for an area-less project. color = #RRGGBB. priority = low\	medium\	high (display-only ! mark, v1.2).
`create_subproject`	[ ]	Create a subproject inside a project (resolved by name or id). priority = low\	medium\	high (display-only ! mark, v1.2).
`create_todo`	[ ]	Create a to-do. Parent (project or subproject) is resolved by exact name (case-insensitive) or id; omit for a standalone to-do. when/deadline accept natural language ("tomorrow", "fri", "7 oct") or yyyy-MM-dd. Pass `repeat` to make it recurring (spec/60 — clears deadline, stamps whenDate/anchor). A repeat rule takes an optional `mode` ("schedule", the default fixed grid, or "afterCompletion", where the next date counts from when you complete it — afterCompletion carries ONLY freq+interval), an optional `until` (strict yyyy-MM-dd, the last date an occurrence may fall on, inclusive) and/or an optional `count` (occurrences remaining incl. the active one, 1…999, decremented on each completion). Setting a repeat with no when in schedule mode seeds the first occurrence date; in afterCompletion mode it seeds when=today. `reminder` is a local `yyyy-MM-ddTHH:mm`.
`delete_container`	[ ]	Alias of trash_container — moves a container subtree to trash (v1.2; recoverable). For an irreversible hard delete use permanent_delete.
`delete_filter`	[ ]	Delete a saved filter (v1.4) — HARD delete, no trash, no confirm (a config row like a tag; cheap to recreate, and the to-dos it matched are untouched). Resolve by id or exact name.
`delete_todo`	[ ]	Alias of trash_todo — moves a to-do to trash (v1.2; recoverable). For an irreversible hard delete use permanent_delete.
`empty_trash`	[ ]	Permanently delete EVERY trashed root and its subtree (v1.2) — the irreversible hard-delete cascade for the whole trash. Destructive — requires confirm: true.
`get_anytime`	[✓]	Anytime (v1.2, narrowed in v1.6): open to-dos with no scheduled date (whenDate null) that ARE filed under a project/subproject. A deadline may exist and the row still lists here (deadline alone never schedules a to-do). someday to-dos are excluded. v1.6: an UNFILED to-do (no parent AND no date) is NOT here — it belongs to the inbox; see get_inbox.
`get_filter`	[✓]	A saved filter’s VIEW (v1.4): the matched to-dos with their count, in the anytime view’s order (tree-position-then-sortKey, standalone last). Universe = open + someday visible to-dos matching ALL criteria; done never matches (the logbook owns history). An INERT filter (criteria from a newer version) returns count 0, no matches, and a muted note. Resolve by id or exact name.
`get_inbox`	[✓]	The inbox = the triage queue (v1.6): UNFILED to-dos first (open, no parent AND no date), then open capture rows in capture order. `openCount` is the nav count over both. A to-do leaves the inbox the moment it gets a parent or a date — there is no inbox flag to set.
`get_logbook`	[✓]	The completed-history logbook (v1.4): every done to-do (non-trashed, with a completedAt), grouped by the LOCAL calendar day it was completed, newest day first, and completedAt-descending within a day. Each day carries its date, a label, a count and the to-dos. Recurring done copies appear like any done to-do; inbox items never appear. Unbounded — no nav count.
`get_next_7_days`	[✓]	The next 7 days (a rolling window: today through today+6): overdue plus one bucket per day, each with its to-dos. Always 7 buckets — it is not bounded by the calendar week and does not read the weekStart setting.
`get_overview`	[✓]	The full task tree: areas → projects → subprojects with recursive open-todo counts and progress, their open to-dos (with dates/tags), standalone to-dos, and an inbox summary. The one call to understand everything.
`get_someday`	[✓]	Someday (v1.2 phase 2 — GROUPED): parked to-dos and parked inbox items. Parking is timeline-free (it strips dates), so no date meta renders here. Top to bottom: `groups` — every project/subproject that has DIRECT parked to-dos, in overview-tree order, each `{ container, todos }`; `standalone` — standalone parked to-dos (no project) then parked TEXT inbox captures; `attachments` — parked url/image/file inbox items, last. `containerSomeday` is the raw `{containerId: [todoIds]}` map that also feeds each container’s own collapsible someday section.
`get_today`	[✓]	Today view: overdue to-dos (missed deadlines), today’s to-dos, and what was completed today.
`get_upcoming`	[✓]	Upcoming: open to-dos dated after today, grouped by their effective date ascending.
`inbox_to_someday`	[ ]	Park an inbox item in someday (stays an inbox item; v1.2 parking strips its captured date — contrast archive, which keeps it).
`list_filters`	[✓]	The saved filters (v1.4), in nav order (sortKey). Each row reports its parsed criteria, a live match COUNT (open + someday visible to-dos matching all criteria; done never matches), a derived priorityMin when the filter leads with a priority criterion, and inert=true for a filter whose criteria this client can’t parse (an unknown/newer-version shape — it matches nothing and renders a muted note). The three defaults low/medium/high are ordinary rows.
`list_trash`	[✓]	The trash view (v1.2): trashed ROOTS only (a row trashed individually inside a container that was later trashed is hidden here but still purges by its own stamp), newest first. Each entry reports its kind, label, when it was trashed, and — for containers — the subtree it will take with it on permanent delete.
`move_container`	[ ]	Move an area/project/subproject to a new parent, and/or reorder it among its new siblings. parent = the destination container (name or id), or the literal "top" for the top level. ⭐ A move can RE-KIND the row, exactly as dragging does: move a project onto a project and it becomes a SUBPROJECT of it; move a subproject to "top" and it is PROMOTED to a project. The hierarchy is closed (area → project → subproject), so illegal moves are refused with the reason — nothing may nest under a subproject, an area is always top level, and nothing may move into its own contents. position = "start" \	"end" (default) \	the id of a sibling to place it after.
`move_inbox_item`	[ ]	Sort an inbox item onto a target (resolved by name or id). A text item MOVES into a container (becomes a to-do, carrying its captured date); a url item ATTACHES to a container or to-do. Pass target "anytime" to turn a text item into a standalone, unscheduled to-do (v1.2).
`move_to_anytime`	[ ]	Move an existing to-do to anytime (v1.2): detach it from its project and clear its scheduled date, keeping its deadline and tags. It becomes a standalone, unscheduled to-do that lists in the anytime view. Resolve by id or exact title.
`move_todo`	[ ]	Reparent and/or reorder a to-do. position = "start" \	"end" \	a sibling to-do id to place it after. Omit `parent` (or pass "none") to reorder within the STANDALONE band — the project-less to-dos that list after the projects on the overview.
`permanent_delete`	[ ]	Permanently delete a trashed row by id (v1.2) — the irreversible hard-delete cascade (a container takes its whole subtree: child containers, to-dos and attachments). Destructive — requires confirm: true. Pass an id from list_trash.
`restore_trashed`	[ ]	Restore a trashed root by id (v1.2): clears its trash stamp; the row — and, for a container, its whole subtree — returns to its views untouched.
`search`	[✓]	Search everything (v1.2): case-insensitive substring match across to-do titles/subtitles/notes/tags, container names/subtitles/notes, inbox text/urls/titles, and attachment titles/urls/text. Excludes trashed rows and archived/trashed subtrees; someday rows are included. Each hit reports which field matched and its direct-parent crumb.
`set_week_start`	[ ]	Set the first day of the week (shared setting): 1 = Monday (the default), 7 = Sunday, null = Monday. v1.5: this is presentation-only — it orders the weekday columns of the date-picker calendar in the apps. It changes NO view, no bucket and no count (next-7-days is always today…today+6), and weekly recurrence stays Monday-anchored.
`trash_container`	[ ]	Move an area/project/subproject (and, by exclusion, its whole subtree) to trash (v1.2). Only this root row is stamped; restore_trashed brings the subtree back untouched. Auto-purged after 30 days. No confirm — trash is the safety net.
`trash_todo`	[ ]	Move a to-do to trash (v1.2). Recoverable via restore_trashed; auto-purged after 30 days. No confirm — trash is the safety net. Resolve by id or exact title.
`uncomplete_todo`	[ ]	Reopen a done to-do (clears completedAt). Resolve by id or exact title.
`update_container`	[ ]	Edit an area/project/subproject. Pass only fields to change. notes = the detail-pane note (v1.2), or null to clear. deadline accepts a date phrase / yyyy-MM-dd, or null to clear (areas have no deadline). priority = low\	medium\	high (display-only ! mark, v1.2; projects/subprojects only — areas never carry it), or null to clear. subtitle/name/color (#RRGGBB) editable too.
`update_filter`	[ ]	Edit a saved filter (v1.4): pass `name` and/or `criteria` (the same { v:1, all:[…] } shape as create_filter — replaces the whole criteria set). An unknown/newer-version criterion is rejected. Resolve by id or exact name. Returns the filter and its new match count.
`update_todo`	[ ]	Edit a to-do. Pass only fields to change. status = open\	someday\	done. Parking (status→someday) is timeline-free for a PLAIN to-do: it STRIPS whenDate + deadline to null but KEEPS the parent (a project idea parks with its project); a RECURRING to-do instead keeps its whenDate + rule DORMANT while parked (spec/60). Un-park a plain parked to-do by setting any when or deadline (assigning only a parent leaves it parked); un-park a recurring parked to-do with status=open, which re-aligns its dormant whenDate to today-or-later (schedule mode) or leaves it (afterCompletion). parent = project/subproject name or id, or null to make standalone. when/deadline = a date phrase / yyyy-MM-dd, or null to clear. priority = low\	medium\	high (display-only ! mark), or null to clear. tags replaces the tag set. repeat = a rule to make it recurring (clears deadline, re-anchors to its whenDate; takes optional mode "schedule"\	"afterCompletion" + optional until yyyy-MM-dd + optional count 1…999), or null to clear the rule. reminder = a local `yyyy-MM-ddTHH:mm`, or null to clear.

What an AI client cannot do

Read anything outside the granted scopes. Scopes are enforced per request.
Write on a read-only grant. The server refuses.
See your data after you revoke. The grant carried the key; revoking ends it.
Search your notes on the server. Everything is decrypted per request, for that request only — see encryption.

Building an MCP client

The server implements the standard discovery documents — RFC 9728 protected-resource metadata and RFC 8414 authorization-server metadata — so a compliant client needs no special-casing.

One thing is stricter than the baseline: resource (RFC 8707) is mandatory on /authorize and /token. A token minted for the MCP server is refused at the REST API and vice versa. Send https://mcp.inittasks.com/mcp as the resource when connecting here.

See authentication for the rest of the flow — it is the same one.