Donut Browser
Donut
Get Started
Skip to content

AI agents

Agent page, recipes and schedules.

View as MarkdownOpen in ChatGPT(opens in a new tab)Open in Claude(opens in a new tab)
On this page

Donut has no AI model. You connect your own agent over remote MCP, and the Agent page (Cmd/Ctrl+J) is where you work with it: you see what it does, answer its questions and step in when a page needs a person. The page keeps this data in memory only; nothing on it is saved to disk.

Before an agent connects

The page shows what is missing: sign in, a plan with remote control (Solo, Pro, Team or Enterprise), or Remote control turned on under Integrations → Remote MCP. When all is set, it waits for your agent and shows Instructions for your agent that you can copy into the agent’s prompt.

Tabs

TabWhat it shows
ConversationQuestions, help requests, notes and progress reports from your agents. Write a note to an agent at the bottom: Enter sends, Shift+Enter starts a new line.
ProfilesThe profiles your agents work in. Take over or Hand back one profile, or several at once.
ActivityEvery tool call your agents made.
RecipesSaved flows an agent can repeat.
SchedulesRecipes that run on a calendar.

The header has the Pause agents switch and Connect an agent.

Agent page, Conversation tab, with an AI agent connected over remote control. The agent reports progress on the Shop profiles, asks a person to enter an SMS code in the window of Shop · US-East 02, and asks whether to combine two orders with the choices Combine, Ship separately and Skip for now. A Pause agents switch and a Connect an agent button are at the top.
Watch what your agent does, answer its questions and step in when a page needs a person.
Agent page, Activity tab. A list of the agent's tool calls, each with the time, the agent (Claude Code), the tool such as list_profiles, list_proxies or get_profile, the profile it touched, the duration in ms and the result OK. Filters: all agents, search by profile, errors only.
Every tool call an agent makes, with profile, result and duration.

Questions and help requests

An agent uses these MCP tools to keep you in the loop:

  • ask_human shows a question. Type the answer in Your answer.
  • request_human_help asks you to act in a profile, for example to sign in, solve a challenge or pay.
  • wait_for_human waits until you answer or hand a profile back.
  • get_human_updates reads the notes you sent.
  • report_progress posts progress to the conversation.

The app shows a notice when an agent asks something, and the Agent item in the sidebar shows how many requests wait.

Take over and pause

  • Take over a profile to work in it yourself. Add a note for the agent, for example “I am fixing this profile, skip it for now”. Until you click Hand back, tools on that profile answer PROFILE_HELD_BY_HUMAN.
  • Pause agents stops every agent until you click Resume: their tools answer AGENT_PAUSED. Add a note to say why. While paused, agents can still read your notes and ask you questions.

Recipes

A recipe is a flow recorded once. An agent or a schedule can repeat it on other profiles without spending model tokens.

  • Click New recipe, choose a running profile, click Record and do the task in its browser. Click Stop recording. Passwords are never recorded.
  • Or write the steps by hand. Step types: Go to a page, Click, Type, Press a key, Scroll, Wait, Wait for, Extract and Screenshot. A step that acts on an element names it by role and name, or by a CSS selector, never both. A run refuses Press a key and Scroll steps with REPLAY_STEP_UNSUPPORTED, so leave them out of recipes you run.
  • Text, addresses and selectors can hold {{name}} placeholders. run_recipe fills them from its variables, so one recipe serves many accounts.
  • An account holds up to 200 recipes, each with up to 200 steps. Recipes are stored in Donut cloud and sync to all your devices.
Agent page, Recipes tab. A recorder writes down the steps you take in a running profile's browser, and never records passwords. The recipe Check new orders is open in the editor with its name and steps: Go to a page https://seller.example.com/orders, Wait for the table named Orders, Click, and more, each with move up, move down and delete buttons.
Recipes are saved step lists an agent or a schedule can repeat.

Running a recipe on demand needs Pro, Team or Enterprise. Solo can save recipes and run them on a schedule on the remote fleet. On Solo, the recipe tools do not appear over MCP.

Schedules

A schedule runs one recipe on a set of profiles. Schedules are part of Solo, Pro, Team and Enterprise.

  1. On the Schedules tab, click New schedule. Pick a recipe and the profiles.
  2. Under Where it runs, pick Remote fleet or This computer. This computer needs Pro, Team or Enterprise, and the app must be open. The fleet spends remote hours and needs profiles that can run remotely.
  3. Under How often, pick On set days (days and a time), Once (a date and time) or At an interval (every 15 minutes to every 7 days). Times use your time zone.

Donut cloud starts each run on time. Run now starts one at once, the switch pauses a schedule, and Runs lists past runs with their results. Under Notices, a Webhook URL (https) gets a signed message when a schedule fails and when it works again. A schedule that keeps failing sends at most one reminder a day. Send a test notice checks it, and New secret replaces the signing secret.

Agent page, Schedules tab with three schedules: Orders at 8:00 runs the recipe Check new orders Monday to Friday at 08:00 on this computer, Warm-up every 6 hours runs Morning warm-up every 360 minutes on the remote fleet, and Support replies is paused. Each row has an on/off switch and Run now, Runs, Edit and Delete. A New schedule button is at the top right.
Run recipes on a calendar, on this desktop or on the remote fleet.

Checked against the app on .