← All MCP tools

Create task

create_task
Write

What it does

Create a task. Break work into subtasks via parent_id, or file a follow-up. Set assignee='claude' to add it to the AI agent queue. Optional: project_id, due_date (YYYY-MM-DD), due_time (local "HH:MM" on that date — a task with one is scheduled AT that time rather than fitted into a free slot), estimate_minutes, acceptance_criteria (what "done" means — the bar the result is checked against), description (long-form write-up, HTML), references (URLs/paths for context — a Google Doc, a repo file — fetch them yourself via whatever connectors you have; Done just stores the pointer), horizon. A task you create starts gated: close it with propose_completion (only a human can mark a task agent_ok). Not settable here, because they are the human's to grant: notes_for_agent (the human's one-way instructions — put your own context in notes or description) and budget_tool_calls/budget_minutes.

Parameters

NameTypeRequiredDescription
titlestringRequired
notesstringOptional
acceptance_criteriastringOptional

What "done" means for this task — the checkable bar the result is judged against, not instructions. Write it as conditions someone can verify ("tests pass on the branch tip", "page loads under 200ms"). Read it before you start and check your work against it before proposing completion.

descriptionstringOptional

Long-form write-up for the task, as HTML — this is what the rich-text editor reads and writes, so send HTML, not markdown. Send the tags raw — do not HTML-escape them (<p>x</p>, never &lt;p&gt;x&lt;/p&gt;), or the write-up renders as literal text. Supported: <h1>-<h3>, <p>, <strong>, <em>, <ul>/<ol>/<li>, <a>, <code>, <pre>, <blockquote>, <table>/<tr>/<th>/<td>, and checklists as <ul data-type="taskList"><li data-type="taskItem" data-checked="false">. Use notes instead for a short one-liner.

prioritystringOptional
Allowed valuesurgentmediumlow
parent_idstringOptional
project_idstringOptional
goal_idstringOptional

File this task under a goal (North Star) as an action step toward it. Get ids from list_goals.

start_datestringOptional

YYYY-MM-DD. Optional: the day work is meant to begin. With due_date it is the task's span on the human's Gantt chart; nothing schedules or reminds off it.

due_datestringOptional

YYYY-MM-DD

due_timestringOptional

Optional local wall-clock time on due_date, 24-hour "HH:MM". A task with one is scheduled AT that time instead of being packed into a free slot.

due_hardbooleanOptional

true = hard deadline (must finish on/before due_date); false (default) = soft, a suggested date

waitingbooleanOptional

true = parked, waiting on someone else. A parked task stays on the board under "Waiting" but leaves the human's calendar feed, Plan my day and Today until its resurface_date arrives (or indefinitely, with none) — so it stops being time-blocked into every day while the ball is in someone else's court.

resurface_datestringOptional

YYYY-MM-DD. The day a parked (waiting) task comes back onto the human's calendar and into Today. Only meaningful alongside waiting=true.

assigneestring | nullOptional
Allowed valuesclaudenull
agentstringOptional

Only on an account with an agent roster (list_agents): pins the task to a named agent. Must be your OWN lane — a task you create lands in your queue, never another agent's; use handoff_task to pass work on. Implies assignee=claude. Without it, an assignee=claude task you create is pinned to your lane anyway.

estimate_minutesnumberOptional

Task length in minutes

horizonstring | nullOptional
Allowed valuestodaythis_weekthis_monthbacklogparking_lotnull
impactnumberOptional

1–5 rating of how much this moves the needle (feeds the Value/priority ranking)

valuenumberOptional

Optional payout/value of this task (dollars, points, etc.); boosts its priority ranking

referencesstring[]Optional

URLs/paths the agent should read for context

deliverablesobject[]Optional

Structured outputs you produced — {label, url, kind, content}. content is the document itself so a human can read and approve it without leaving Done; url is for something that genuinely lives elsewhere. kind says what the document is — markdown (the default), html, or pdf as base64. The output-side counterpart to references.

deliverables[].labelstringRequired
deliverables[].urlstring | nullOptional
deliverables[].kindstringOptional

What this deliverable holds. Omit for markdown, which is the default and the right answer for a draft, summary, report or plan. Use "html" only for a page you actually generated as HTML — it is rendered in a locked-down sandbox that cannot run scripts or reach the network, so a page that depends on either will look wrong. Use "pdf" with base64 in content for a small PDF, or with artifact_id for a large one. Use "binary" for anything else you uploaded — a screenshot, a spreadsheet, a zip — which is always an artifact_id and has no inline form.

Allowed valuesmarkdownhtmlpdfbinary
deliverables[].contentstring | nullOptional

The document itself, so a human can read it without leaving Done. Markdown or HTML as text, capped at 10,000 characters each (40,000 across the task). A pdf goes here as base64, capped at 800,000 base64 characters (about 600 kB, and 1,200,000 across the task). Put anything larger behind a url rather than truncating it.

deliverables[].artifact_idstring | nullOptional

An artifact you uploaded, instead of inline content (the two are mutually exclusive). Upload first: call create_artifact_upload with {task_id, label, filename, kind, size_bytes}, then PUT the raw bytes to the one-time upload_url it returns — no Authorization header, the link itself is the credential, and it expires 15 minutes after you request it. Use the artifact_id it returned here. To revise the document later, call create_artifact_upload again with the SAME artifact_id — this pointer keeps working and starts resolving to the new file as soon as its bytes land. Done keeps only that latest file: the previous one is queued for deletion once the replacement is safely stored, so don't rely on being able to fetch an earlier upload back.

deliverables[].filenamestring | nullOptional

What the downloaded file should be called. Ignored for the text kinds, and sanitised before use — the extension comes from an allow-list, not from you.

idempotency_keystringOptional

Optional. One key per operation: pass the same key only when retrying this exact call, so a retry after a timeout replays the first result (a key already used on a different tool is refused, and keys expire after 48 hours) instead of creating a duplicate task.

Annotations

The hints the server sends with this tool. A host reads them to decide whether to ask you before a call.

Read-onlyreadOnlyHint: false
Changes data in your Done account.
DestructivedestructiveHint: false
Only adds or reads; nothing already there is removed or overwritten.
IdempotentidempotentHint: false
Each call can add something new, such as a comment, a run or a task.
Open worldopenWorldHint: false
Touches only your own Done account.

Which tokens can call it

A connection made with OAuth has full access. A token from Settings → Agent access → Extra MCP tokens can be narrowed, and a narrowed token never sees beyond its project or workspace.

  • Full accessAllowed
  • Read-only tokenRefused
  • Single-project tokenAllowed
  • Single-workspace tokenAllowed

Try it

With Done connected, ask your assistant something like:

“Add an urgent task to renew the domain by Friday, and assign it to Claude.”

Connect Done to your assistant

One MCP endpoint per account, over OAuth or a token. Step-by-step setup for each client:

MCP setup overview →