Workflows, schedules and alerts

​

Save a workflow

• Under any of Mobster’s answers in a conversation that went well, choose Save as workflow. A new workflow’s page opens with the task’s own words while Mobster reads the conversation and drafts the name and the steps, with the parts you’d change marked as blanks; the draft takes their place unless you’ve changed them first. Nothing is saved until you choose Save workflow. With several phones it runs on the one the task ran on, and Always continue in the same conversation adds every run to that conversation.
• On the Workflows page, choose New workflow and write the steps yourself, on the workflow’s own page (The page).
• On an empty Workflows page, choose one of the ideas under Start from one of these. Each shows the icon of the app it works in (Calendar, Mail, Instagram, Amazon, Settings), and an idea that spans no single app shows a plain glyph. It opens a new workflow’s page filled in, schedule included, and saves nothing until you do.

​

Blanks

​

Run a workflow

• Workflows page: Run. A workflow with blanks asks for them in its row first.
• ⌘K: type its name, then choose Run “…”.
• In a conversation: choose Run workflow here… in the conversation’s ··· menu in the sidebar, or in ⌘K. The steps become that conversation’s next message, so the agent sees everything before them.
• Siri and Shortcuts: mobster-task.sh --workflow "Reels digest" topic=basketball (Siri, Raycast and Alfred).
• Terminal: mobster chat --workflow "Reels digest" --set topic=basketball (mobster chat).
• Your coding agent, over MCP: list_workflows and run_workflow (MCP server).
• The HTTP API: POST /api/workflows/{id}/run with an Idempotency-Key header and {"values": {"topic": "basketball"}, "threadId": "…"}. Both fields are optional. The answer has the run and its threadId.

​

Run it

• On a schedule: turn it on, and pick the days and the time beside its switch (Schedules).
• By webhook: in Edit, once the workflow is saved, Make a URL gives another tool a signed address that starts the workflow (Start a workflow from another tool).
• From a shortcut: Copy the command copies the line for Siri, Shortcuts, Raycast or Alfred, with the workflow’s name and a NAME= for each blank without a default (Run a saved workflow).

​

What a workflow’s row shows

• What it does: its name, and its steps’ first line under it.
• Worked: when its last verified run finished, and the first line of what that run found, such as Worked 8:02 AM beside “Standup at 9:30, lunch with Priya at 12”. A run is verified when the screen proved it done or you marked it Looks right. Choose Worked to open that task.
• Next run: when a scheduled workflow runs next. A workflow you run by hand has none.
• What needs you, in place of Worked, from its newest attempt that ended:

​

Learned steps

• A send, post, delete or purchase still asks you first, as in any task. Each replayed step goes through the same rules as a live step: in Ask before every change a Save or Create asks, a like, follow or repost asks, and typing a one-time code asks in every mode. In Full access, Mobster hands that step to the model, which goes ahead as Full access says. A replay never opens or acts in an app you haven’t allowed: it hands that step to the model, which asks before it opens the app.
• Typed text comes from the workflow’s own words: a workflow that says “Text : ” types the message you fill in. A search query replays as it was typed, because it finds things rather than writing them. Text the model wrote, such as a caption, is written by the model each run.
• When a screen doesn’t match (the app changed, a sheet popped up), the model takes over from that step with what was done so far, and the run that finishes is learned again. After a replayed send, post or delete, or when a learned action’s outcome is unknown (the phone didn’t answer, or the screen couldn’t be read after it), the model never takes over: the task ends with Check the result, so nothing is done twice. Look at your iPhone before you run it again.
• A replay says done only on evidence. Either the screen shows what the learned run quoted at the end, such as the sent message, or the replay ends on the learned run’s last screen in the same app with the words it typed on it, such as the new note’s text. When neither holds, one look-only call to your model checks the screen; it can’t tap or type, and its words count only if they’re on the screen. Otherwise the run ends with Check the result. The run’s replay.confirmed says which: proof, screen or look.
• A loop over rows (Loop › Each row) whose first round is learned replays the next rounds with no model, each with its own row’s words, and goes on to the next round when each one ends done.
• A workflow that reads something (an answer schema, or only taps and scrolls) replays its steps, and the model reads the answer at the end.

​

Learned steps from any task

• What’s kept: when a task finishes with every part shown on screen, Mobster keeps its steps, its checklist and your request, with the values you typed or named as blanks. “Create a note titled ‘Errands 5’ that says ‘post office’” also serves “Create a note titled ‘Garden 6’ that says ‘water the basil’”. A row Mobster scrolled to because your request named it counts as a value too: “Open Settings › Accessibility › Subtitles & Captioning” also serves “Open Settings › Accessibility › Keyboards & Typing”. Each screen is kept as hashes of its controls’ roles and names, never as text. Answers, notes and anything Mobster wrote itself aren’t kept.
• When it’s used: a later task in the same app, on the same iPhone, whose words are the kept request, with the same values or other ones. Code fills in the values from your words. The model is asked only when your words could fill the request two ways. A task that says more than the kept request, such as a second request after “then” or after a closing quote, runs with the model instead: a value never runs across a quote mark, a semicolon, the end of a sentence, or a “then” outside quotes.
• What’s repeated: each step runs only on a screen like the one it ran on, and types only your words. An app that reopens on another page is first brought back to where the steps started: its Back button, then closing and reopening it once, never while a field on screen holds typed text. Sends, payments and deletes ask you first, as in any task; in Full access they go ahead as Full access says. When a screen doesn’t match, the model takes over with the remaining steps as a plan.
• How it finishes: the first repeat of learned steps ends the usual way, with the model checking the screen. After one repeat that shows every part done, later repeats of a task with nothing to report end as soon as the screen shows it done, with no model.
• What you taught: a task you taught on the live view (Teach Mobster a task) is kept as learned steps too. It opens its own app first, so it runs whichever app a task starts in, and it also runs for a workflow’s runs and its schedule. It’s kept until you forget it: a run that fails leaves it to be checked by the model again. Save as a workflow makes a workflow whose words run it, with your blanks as the workflow’s.
• Forgetting: learned steps whose first repeat fails are forgotten at once. After that, ones that fail twice in a row are forgotten. Mobster keeps up to 100 tasks’ steps, in the data folder’s replays/programs/.
• Parts of learned steps: with MOBSTER_PROGRAMS=on set in the env file, a new task with other words can also use parts of other tasks’ learned steps. A calories question’s steps have a “look up a food in CalTrack” part; a create-a-note task’s steps have a “create a note with this title and text” part. “Create a note titled ‘Groceries 12’ that says ‘rice and beans’, then tell me the calories in one serving of Quinoa in CalTrack” then runs both parts with no model:
  • code takes each value from your words, where the kept request had its value, and checks it against the task’s checklist;
  • a part runs only for a part of the checklist still open in its app, and a part that wrote something never runs for a task that only reads;
  • each step runs only on a screen like the one it ran on, and the part’s last screen must show your values;
  • a value your words don’t give is left to the model, with the remaining steps as a plan.
  On iOS Simulators (12 Oct 2026), 6 code-graded tasks that combine parts of other tasks’ learned steps took 22 model calls instead of 39 and $0.111 instead of $0.198; 6 of 6 passed both ways.

​

The Morning brief

• Three task cards, each read-only: “Don’t send anything”, “Don’t like, comment or follow”, “Don’t like, comment or save anything”. A reply goes only when you approve its draft, and that send asks as any send does.
• A Brief step that puts their results together (Build a workflow on its canvas).
• Every day at the time you chose, with a $2.00 spending limit a run and a 60-minute time limit.

​

Schedules

​

When a schedule runs

• A step that needs your approval waits for you, as in any task. If nobody answers within 10 minutes, the task ends without that step.
• A workflow saved with a profile (account in the API) runs on that profile’s phone, as that profile, with its values in the task’s placeholders (Profiles).
• A workflow runs on the phone saved with it. With none, it runs on the only phone, or the first iPhone, never on whichever phone you used last.
• Open at login, in Settings › General, starts Mobster when you log in, so schedules run even if you haven’t opened it. Closing the window keeps Mobster running; ⌘Q quits it.
• Task notifications, in the same place, tell you when a task finishes or needs your OK while Mobster is in the background.

​

Staying awake

​

Alerts

MOBSTER_ALERT_WEBHOOK_URL=https://hooks.slack.com/services/…

• Slack, Discord and ntfy addresses get their own format. Any other address gets JSON: workflow, status, reason, runId, at and text.
• A post says which task and why. It never holds the task’s answer, a screenshot or anything it typed.
• Tasks you start yourself never alert.
• Addresses on this Mac or its local network are refused.