Scheduler

Mobile Agent can run tasks on a schedule — hourly digests, daily reports, nightly file cleanup, weekly summaries — while the app is closed. Jobs are driven by a background service with exact alarms and can write results back to the chat library.

Create a schedule

1. Open Settings → Jobs (or the scheduler section) and tap New schedule.
2. Enter a title and the prompt the agent should run (e.g. "Summarize today's notes in the workspace and save to daily.md.").
3. Pick a frequency:
   • hourly, daily, weekly, monthly — preset intervals.
   • custom — any cron expression (via the croner engine), e.g. 0 9 * * 1-5 for weekdays at 9:00.
4. Select the model (provider + model), and optionally an agent (see Custom Agents).
5. Toggle autoApprove — with it off, tool approval requests are raised; see Tool Approvals.
6. Optionally attach an external folder session (see File Access) so the job can work in a folder you mounted.
7. Save. The scheduler shows nextRunAt and lastRunAt, and you can enable/disable the job anytime.

How runs work

• An in-app scheduler engine ticks every 30 seconds and fires due jobs.
• Alarms use the native exact alarm API (SchedulerAlarm) so a job fires close to its scheduled time even when the app is backgrounded.
• A foreground service (BackgroundAgentService) keeps runs alive, shows a persistent notification, and was designed to work with the battery-optimization exemption (the app asks you to allow it on first schedule).
• Each run is recorded with a status: queued → running → completed | failed | skipped.
• Files created during a run are saved and restored on wake, so a job that writes a report before the next alarm keeps its outputs.

Notifications

You choose which events notify you in Settings → Notifications:

• Approval requests — when a non-auto-approved tool needs your input.
• Run finished — when a scheduled run completes or fails.

Because jobs run in the background, notifications are the primary way you learn about results without keeping the app open.

Timezones and timing

Every schedule stores its own timezone, so cron expressions are evaluated in the timezone you set, not the device's. If you travel, the job keeps its meaning.

Jobs also have a 60-minute grace window for missed or delayed firing — if a run is missed (e.g. device was off), the scheduler attempts it within the window rather than silently dropping it.

Example

Title: Morning standup prep
Prompt: Open the workspace, read yesterday's `notes.md`,
        and write today's standup draft to `standup.md`.
Frequency: daily at 08:30  (cron: 30 8 * * *)
Model: OpenAI gpt-4.1-mini

The job runs every weekday morning, produces standup.md, and notifies you when done — no app interaction needed.

Requirements

• Battery optimization exempted for the app (prompted automatically) — otherwise the OS may kill the background process.
• Notifications allowed if you want run results.
• A model that supports tool use if the prompt needs tools — on-device reasoning models work fully offline for scheduled tasks.