# Done

> Done is an AI to-do app and task manager. Capture tasks in plain language, then
> assign the ones you don't want to do yourself to an AI agent — Claude, ChatGPT, or
> any MCP-compatible model — that reads the task's context, does the work, and
> reports back in the task's activity feed.
> Every account gets its own remote MCP endpoint, scoped to just that account's
> tasks: there is no shared or global token.

Live at https://trydone.app. Sign up with email + password. The free tier includes
every core feature and 100 AI agent (MCP) calls a month, no credit card required;
Pro ($5/mo) adds unlimited agent calls, and Teams ($20/mo) adds shared workspaces.

## Core concepts

- Tasks have a status (not_started, in_progress, plan_proposed, pending_approval,
  done), a priority, an optional project, goal, and tags, an assignee (a person or
  an AI agent), due dates, recurrence, and "blocked by" dependencies.
- Assigning a task to an AI agent is delegation, not hand-off: a human stays the
  accountable owner and reviews plans and results in the task's Activity panel.
- The completion approval gate: a task with `agent_ok=false` can only reach `done`
  through a human approval (`approve_task`), never directly from an agent. A task
  with `agent_ok=true` can be closed directly by the agent via `complete_task`.
- The plan approval gate: before substantial or ambiguous work, an agent can
  `propose_plan` — the task moves to `plan_proposed` and leaves its ready queue
  until a human approves it (back to `in_progress`, work begins) or rejects it
  (back to `not_started`, with feedback).
- Goals are the North Star tier above projects — file projects/tasks under a goal
  to track progress toward it; deleting a goal unfiles its work rather than
  destroying it.
- Deleting a task moves it to Trash (30-day auto-purge with restore), not a hard
  delete.
- Acceptance criteria are a checklist an agent ticks off as it satisfies them
  (`check_criterion`) — a progress channel distinct from narrating status into
  comments. Ticking every box does not complete the task on its own.
- An agent that hits a real fork — spend money, anything outward-facing — parks
  the question instead of stalling (`ask_decision`, with a recommendation) and
  moves to the next task. The human answers in one click (or from a Slack DM);
  the task comes back with the answer attached.
- Unattended overnight runs are capped per session (tasks, minutes, tool calls),
  independent of any one task's own budget, so a scheduled run can't spend a
  week's headroom in one night (`check_night_budget`).

## Connecting an AI agent (MCP)

1. Generate a per-account MCP token from the Account panel in the app (shown once,
   at generation).
2. Point any MCP-compatible client — Claude, ChatGPT, Cursor, or a custom agent —
   at `https://trydone.app/mcp` with `Authorization: Bearer <token>`. Clients that
   speak OAuth (claude.ai, ChatGPT connectors, Claude Desktop) can instead add Done
   as a custom connector by URL and authorize themselves — no token to copy.
3. Full setup: https://trydone.app/docs/mcp

Done is a full OAuth 2.1 authorization server for its MCP endpoint, and also accepts
long-lived per-account bearer tokens. Either way, auth is per account: tokens and
grants are never shared or global, and only ever see their own account's data.

## MCP tools (workflow-shaped, not raw CRUD)

`list_my_tasks`, `next_task`, `list_tasks`, `get_task`, `claim_task`, `release_task`, `decline_task`,
`comment_on_task`, `delete_comment`, `check_criterion`, `ask_decision`,
`list_open_decisions`, `withdraw_decision`, `check_night_budget`, `complete_task`,
`propose_completion`, `approve_task`, `reject_task`, `propose_plan`, `approve_plan`,
`reject_plan`, `create_task`, `update_task`, `delete_task`, `quick_capture`,
`decompose_task`, `render_plan`, `set_step_status`, `block_task`, `unblock_task`,
`start_run`, `end_run`, `check_budget`, `get_digest`, `get_weekly_review`, `get_stats`,
`list_claude_shipped`, `list_stale_tasks`, `get_evidence`, `get_autonomy_ledger`,
`propose_edit`, `propose_grooming`, `list_grooming_proposals`, `propose_capture`, `list_capture_inbox`,
`fleet_status`, `list_triggers`, `claim_trigger`, `search_activity`, `changes_since`, `list_projects`, `create_project`,
`update_project`, `delete_project`, `list_workspaces`, `list_goals`, `create_goal`, `update_goal`, `delete_goal`,
`list_tags`, `create_tag`, `delete_tag`, `tag_task`, `untag_task`, `list_trash`, `restore_task`,
`list_calendar_events`, `suggest_event_matches`, `link_task_to_event`, `unlink_task_from_event`.

A typical agent run: `get_digest` → `list_open_decisions` → `next_task` (or
`list_my_tasks` outside a fleet) → `get_task` (which also returns the project's standing
brief for agents) → work, ticking `check_criterion` as each condition is met →
`comment_on_task` once with the write-up → `end_run` with receipts (the commit, the CI
run, the test tally — evidence a human can check instead of taking the summary's word)
→ `complete_task` (or `propose_completion` if the task isn't `agent_ok`).

Two workflows worth knowing about. **Grooming:** `list_stale_tasks` returns the open work
that has gone quiet, with the evidence for judging it and a reason code per task; the
agent proposes a disposition via `ask_decision`, a `resurface_date` or `decompose_task`.
Only a human can set `wont_do`. **Approving:** `get_evidence` gathers criteria progress,
spend against budget, receipts, deliverables and trace links into one read, and flags a
run that left nothing checkable behind.

## Docs

- Getting started: https://trydone.app/docs/getting-started
- MCP setup: https://trydone.app/docs/mcp
- REST API: https://trydone.app/docs/api
- Pricing: https://trydone.app/pricing
- Comparisons vs other to-do apps: https://trydone.app/compare

## Not built yet, on purpose

Polling (not push/SSE) for UI updates; drag-and-drop.