Why your AI assistant should see the screenshot, not just the task title
Most tasks are written as a note to yourself: "fix the checkout bug", "update the hero", "reply to Dana". The title is enough for you, because you remember the screenshot you took when you wrote it. An AI assistant connected to your task list gets only the title, so it can reorder and rename but cannot tell what the work actually is.
GetItDone is built so that the assistant sees what you see. This post walks through how, and through the one place where giving an assistant write access to a task list gets tricky: repeating tasks.
Attachments are readable, not just listed
When you attach a screenshot, a design export or a document to a task, the MCP server does not just report that a file exists. The get_task_details tool returns the full content of a single task: its rich-text notes, every attachment (image, video or document) with short-lived URLs the assistant can fetch to read the actual content, and the dependency chain showing which tasks block it and which it blocks.
In practice that means you can ask Claude, ChatGPT or any MCP client "what is blocking the checkout fix?" and the answer can come from the screenshot of the error you attached, not from guessing at the title.
Screen recordings become frames
A screen recording is a natural way to explain a bug, and it is awkward to hand to a language model. The get_task_video_context tool extracts a few evenly spaced frames from any uploaded video in the task notes or attachments, and surfaces any YouTube or Vimeo links found in the notes. The assistant reads those frames as images.
It is deliberately a separate tool: the assistant asks for video context when a task references a recording, instead of paying for frame extraction on every task it lists.
Repeating tasks need a per-day record
The second thing a plain task list gets wrong for an assistant is repetition. If "water the plants" repeats every day and a single status field says "done", then either every future Tuesday starts out ticked, or an assistant that edits the task can quietly flip yesterday back to not done.
GetItDone keeps recurrence at the level of the task itself and records each day separately in an occurrence ledger. Completing Monday does not touch Tuesday, and an edit to the task does not rewrite which days were completed. The complete_task_occurrence tool is the assistant-facing side of this: it marks one day of a repeating task done, and that day can be in the past ("I did it yesterday") or a scheduled future day finished early, without breaking the chain.
Because every surface writes through the same operations, the board checkbox, the REST API and the MCP tools all agree on which days are done.
What this does not do
Your assistant only sees what you attached or wrote in the task, and only through a connection you authorised. Reading and writing are separate scopes (tasks:read and tasks:write), so you can connect an assistant that can look but not change anything. Capturing tasks from a screenshot happens inside the app and always asks you to approve each proposed task before anything is created.
Try it
Connect an assistant by following the MCP guide at /docs/connecting-ai-assistants, then attach a screenshot to any task and ask about it.