Skip to content

myurasov/Superagent

Repository files navigation

███████╗██╗   ██╗██████╗ ███████╗██████╗  █████╗  ██████╗ ███████╗███╗   ██╗████████╗
██╔════╝██║   ██║██╔══██╗██╔════╝██╔══██╗██╔══██╗██╔════╝ ██╔════╝████╗  ██║╚══██╔══╝
███████╗██║   ██║██████╔╝█████╗  ██████╔╝███████║██║  ███╗█████╗  ██╔██╗ ██║   ██║
╚════██║██║   ██║██╔═══╝ ██╔══╝  ██╔══██╗██╔══██║██║   ██║██╔══╝  ██║╚██╗██║   ██║
███████║╚██████╔╝██║     ███████╗██║  ██║██║  ██║╚██████╔╝███████╗██║ ╚████║   ██║
╚══════╝ ╚═════╝ ╚═╝     ╚══════╝╚═╝  ╚═╝╚═╝  ╚═╝ ╚═════╝ ╚══════╝╚═╝  ╚═══╝   ╚═╝

      your personal-life chief of staff  ·  local  ·  private  ·  AI-native

Bills, health, home, family, finances, vehicles, pets, hobbies, important dates. Captured ambiently. Surfaced when it matters. Never sent anywhere.


Contents


What it is

Superagent turns Cursor's AI assistant into the chief of staff a wealthy person would hire — minus the salary, the trust trade-off, and the cloud account. It runs as a folder you open in Cursor: a vault for everything that runs your life, plus the discipline to surface what needs your attention before you have to ask.

Quick-start works in 5 minutes with zero data-source setup. Heavy ingestion (years of email, banks, health, smart home) is opt-in, deferred, and reversible.

What you'll feel

  • Day 1 — a clean folder gets scaffolded. You log a bill. The agent surfaces it on the right day.
  • Week 1 — you've added 5–10 things by saying them in chat. The morning briefing tells you what's due today, what's tomorrow, what's slipping.
  • Month 1 — ingestion is on for email + calendar. The agent catches the trial about to convert, the bill 30% higher than usual, the dentist appointment your daughter just confirmed by email.
  • Year 1"When did I last see Dr. Smith?" answered in milliseconds. "How much have I spent on streaming this year?" categorized and surfaced. "Draft the executor packet" — one document an executor could use to take over your affairs.

The bet: AI is finally good enough to take the administrative load of modern life off your shoulders, surgically and ambiently, so the work that matters has more room.

Get started in 5 minutes

  1. Open this folder in Cursor.

  2. In chat:

    Follow AGENTS.md and run init.
    
  3. The agent reads AGENTS.md, finds superagent/skills/init.md, asks 3 orientation questions (your name, household, what hurts most today), scaffolds workspace/, and walks through one capture skill that matches your top pain point.

After init, the five commands you'll use most:

whatsup            30-second status check
daily-update       full morning briefing
weekly-review      Friday/Sunday wrap
add-bill           capture a recurring bill
add-appointment    capture an upcoming appointment

Or just describe what you want in plain English — the agent matches your phrasing against the skill catalog.

What it can do

Capabilities grouped by intent. Full skill list (~50) lives at superagent/skills/_manifest.yaml.

Intent Skills Example use
Capture (say it; the agent files it) add-bill, add-subscription, add-appointment, add-important-date, add-contact, add-account, add-asset, add-document, add-domain, add-project, add-source, log-event, health-log, vehicle-log, home-maintenance, pet-care, inbox-triage "Just signed up for Spotify Family, $17/mo."
Recall (ask; the agent answers from local data first) world graph, per-domain history.md, audit, events "Show me everything connected to my mechanic."
Surface (the agent reaches out before you have to ask) whatsup, daily-update, weekly-review, monthly-review, appointments, bills, subscriptions, important-dates, follow-up, triage-overdue, personal-signals The morning briefing tells you the trial converts in 3 days.
Plan (time-bounded efforts + scenarios) projects, pm-review, play (playbooks), scenarios, handoff "Plan the kitchen renovation. Budget $25k. Done by August."
Self-improve (the framework improves itself) supertailor-review, doctor Every 90 days: ranked framework-improvement suggestions, with a hard safeguard against personal-data leakage into committed code.

Data sources

Quick-start works with zero ingestion. Catalog of 27 supported sources, opt-in:

  • Email + calendar — Gmail, iCloud, Outlook, Google Calendar
  • Reminders + notes — Apple Reminders / Notes, Obsidian, Notion
  • Finance — Plaid, Monarch, YNAB, generic CSV
  • Health + wearables — Apple Health, WHOOP, Strava, Garmin, Oura, Fitbit
  • Smart home + vehicles — Home Assistant, SmartThings, Tesla
  • Communications — iMessage, Slack
  • Files + media + location — photos via exiftool, Google Maps Timeline

Per-source install / probe / writes destinations / caveats: superagent/docs/data-sources.md.

Privacy

  • workspace/ is gitignored. Local-only to your machine unless you copy it somewhere.
  • No telemetry. No metrics, no crash reports, no "anonymous usage data". Ever.
  • No remote write by default. Ingestors are read-only.
  • Hard safeguard in the framework's self-improvement loop — token-scan that prevents your personal data from leaking into committed framework code, even if you ask it to.
  • Credentials never stored in plaintext. Each account row carries a vault_ref pointing at your password manager.

The long version: superagent/docs/faq.md and superagent/contracts/privacy.md.

Deeper docs

Doc What's in it
AGENTS.md Canonical operating rules — what the agent reads on every Superagent turn
superagent/docs/architecture.md Mental model, repo layout, dual-agent loop, current build status
superagent/docs/data-sources.md Per-source install + probe + caveats (27 sources cataloged)
superagent/docs/domain-guide.md Per-domain practical guide
superagent/docs/faq.md Naming, comparisons, security, multi-user, mobile, what-ifs
superagent/docs/roadmap.md T-shirt-sized backlog, re-prioritized continuously by the Supertailor

Roadmap

T-shirt-sized (XS/S/M/L/XL) with rationale and "done when" criteria. Headline near-term work: implement the highest-leverage ingestors (gmail, google_calendar, apple_health, plaid) and wire the auto-capture rules they enable. Full plan: superagent/docs/roadmap.md.

Contributing

If you're using this and find friction, the most useful thing you can do is tell the agent. Action signals get captured into _memory/action-signals.yaml, the Supertailor digests them, and approved fixes ship as code.

If you want to write code: superagent/supercoder.agent.md documents the conventions. Add a new ingestor by dropping a file in superagent/tools/ingest/<source>.py that subclasses IngestorBase, registering it in _registry.py, and adding a smoke test.

License

Apache License 2.0 — copyright 2026 Mikhail Yurasov. The framework code is open-source under Apache 2.0; your workspace/ data is yours alone (gitignored, never published, never licensed by this project).

About

Let AI do boooring stuff in your life

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors