Memory, Gmail, projects, learning and daily plans in one private workspace.
AiOS is the personal assistant layer for What Do You Do and the wider AiOS idea: it runs on your machine, keeps data local, and turns scattered signals into a clean daily workspace.
| Overview | Opportunities |
|---|---|
 |
 |
| Memory | Profile |
|---|---|
 |
 |
| Planner | Mobile companion |
|---|---|
 |
 |
| App | Job |
|---|---|
| What Do You Do | Observes activity and answers where the day went. |
| AiOS Assistant | Connects mail, projects and memory, then decides what should happen next. |
They communicate only through loopback APIs. Raw activity and email content stay on the device.
- Remembers projects, goals, learning paths, notes, and next actions.
- Tracks hackathons, job updates, Gmail signals, reminders, and wellbeing events.
- Owns Email Intelligence: multi-account Gmail OAuth, encrypted local tokens, local email sync, AI understanding, semantic search, and daily/weekly planning.
- Turns hackathons, repos, email tasks, learning videos, and goals into one planning board with work done, work left, deadlines, and next questions.
- Runs local desktop automation previews before touching files.
- Plans goals into daily, weekly, or monthly roadmaps.
- Connects with Gmail OAuth and local import folders.
- Keeps AI local-first with Ollama support and rule-based fallback.
- Ships as an installable Windows desktop app plus Arch/Linux packaging.
flowchart LR
A["Gmail / files / browser / wellbeing"] --> B["Local Flask API"]
B --> C["SQLite memory + activity store"]
B --> D["Rule engine or Ollama"]
C --> E["Desktop dashboard"]
D --> E
E --> F["Plans, reminders, pipelines, reports"]
AiOS is the integration and planning brain for WDYD v2. It owns Google OAuth, Gmail sync, token encryption, local email storage, Ollama analysis, daily/weekly plans, follow-up suggestions, and semantic search.
Local APIs exposed to WDYD and other loopback clients:
GET /api/intelligence/accounts
POST /api/intelligence/accounts/google/connect
GET /api/oauth/google/sign-in/<job-id>
POST /api/oauth/google/sign-in/<job-id>/continue
POST /api/oauth/google/sign-in/<job-id>/cancel
PATCH/DELETE /api/intelligence/accounts/<id>
POST /api/intelligence/accounts/<id>/sync
POST /api/intelligence/sync
GET /api/intelligence/today
POST /api/intelligence/daily-plan
POST /api/intelligence/weekly-plan
GET /api/intelligence/search?q=internship
GET /api/planning-events
POST /api/planning-events
PATCH /api/planning-events/<id>
Email content is never sent to cloud AI providers by this module. Analysis uses Ollama when available and a deterministic local fallback otherwise.
Google sign-in runs as a cancellable background job. AiOS stays responsive,
shows the browser handoff state, and returns HTTP 202 plus polling URLs to API
clients instead of blocking a request while the OAuth browser is open.
The command planner is the bridge for real-life planning. It creates one row per event, keeps progress notes local, preserves your manual updates across refreshes, refreshes linked repo activity when possible, and exposes today/week/month agenda summaries plus timed plan blocks for WDYD.
When email analysis extracts a deadline such as today, tomorrow, by Friday, or a numeric date, the generated email task carries that due date into the command planner row.
For private GitHub repos or fewer rate-limit headaches, set GITHUB_TOKEN in Settings. AiOS uses it locally only when fetching latest linked repo activity.
Use smaller or quantized local models first:
ollama pull qwen2.5:3b
ollama pull llama3.2:3b
ollama pull gemma2:2bRecommended settings:
AI_PROVIDER=ollama
OLLAMA_URL=http://localhost:11434
OLLAMA_MODEL=qwen2.5:3b
OLLAMA_EMBED_MODEL=nomic-embed-text
The rule-based fallback still extracts urgent emails, deadlines, and action items when Ollama is offline.
Use Settings -> Test Ollama to check the local Ollama server and whether the selected model is installed. AiOS refuses non-loopback Ollama URLs so email/planner content stays local.
Manage Gmail from Settings -> Google account:
- select Sign in with Google to connect the first account
- select Add another Google account for additional accounts
- rename an account
- pause or resume sync
- sync one account or all accounts
- remove an account, revoke Google access, and delete its local token
The installed app includes its Google desktop client configuration. Users never paste keys or import JSON. The browser flow uses PKCE, a random loopback port, account selection, and read-only Gmail access. See Gmail OAuth for privacy and release-maintainer notes.
Set EMAIL_SYNC_INTERVAL_MINUTES in Settings to control continuous background sync. The worker enforces a 2-minute minimum to avoid hammering Gmail.
flowchart TB
S["Shared sidebar shell"] --> O["Overview tabs"]
S --> M["Memory"]
S --> P["Planner"]
S --> A["Automation"]
S --> B["Browser Agent"]
S --> C["Career Copilot"]
S --> X["Sources / Connectors / Workers / Settings"]
O --> T["Overview | Opportunities | Reminders | Inbox AI"]
Everything important now uses the same shell, top pipeline rail, profile button, and smooth page motion.
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
copy .env.example .env
python run.pyOpen:
http://127.0.0.1:5000
Run the packaged desktop build:
.\release\AiOS-Assistant.exeBuild and install the desktop app:
.\scripts\build-desktop.ps1
.\scripts\install-desktop.ps1 -EnableStartupThe installer copies AiOS-Assistant.exe to %LOCALAPPDATA%\Programs\AiOS Assistant, adds Start Menu/Desktop shortcuts, and can enable login startup. The startup launcher opens AiOS in background tray mode; closing the desktop window hides it to the tray until you use Exit AiOS from Settings or the tray menu. When the desktop app starts, it owns the background loops for reminders, imports, opportunities, and activity tracking.
Arch/Linux:
./scripts/build-desktop-arch.sh
tar -xzf release/AiOS-Assistant-arch-x86_64.tar.gz -C /tmp/aios
/tmp/aios/install-arch.sh --enable-startupAiOS works without a model by using deterministic local rules. For local LLM planning/classification:
ollama pull qwen2.5:3b
ollama pull nomic-embed-textSet in .env:
AI_PROVIDER=ollama
OLLAMA_MODEL=qwen2.5:3b
OLLAMA_EMBED_MODEL=nomic-embed-textapp/
routes.py Desktop pages, API endpoints, OAuth routes
models.py SQLite models
services/ Memory, planner, email intelligence, connectors, workers, settings
templates/ Shared desktop shell and pages
static/ CSS, JS, manifest, icons
automation_agent/ Local file and office automation tools
browser_agent/ Browser research and job tracking planner
career_agent/ GitHub, resume, roadmap, and job match logic
docs/ Architecture, QA, screenshots, module specs
extension/ Browser/plugin companion surface
packaging/ Desktop release helpers
tests/ Regression and integration tests
| Page | Purpose |
|---|---|
/ |
Desktop overview with tabbed Overview, Opportunities, Reminders, Inbox AI |
/gmail |
Gmail intelligence feed |
/hackathons |
Hackathon corner |
/jobs |
Placement and job tracker |
/wellbeing |
What Do You Do / activity signals |
/memory |
Persistent personal memory |
/planner |
Goal planner |
/automation |
Desktop automation preview and approval |
/browser-agent |
Browser research and job search agent |
/career |
Career Copilot |
/profile |
Name, role, current focus, profile photo |
/connectors |
Gmail and import connectors |
/workers |
Desktop background service status |
/settings |
Local config, PIN lock, desktop startup |
- Credentials stay out of git:
credentials/,.env,instance/,release/,dist/, andbuild/are ignored. - Gmail tokens live locally.
- OAuth refresh tokens are encrypted before storage.
- Destructive automation uses previews and approval.
- Local API pairing is loopback-only and token protected.
- Cloud AI is optional; local-first is the default design.
python -m pytest -q
python -m pip_audit -r requirements.txt
python -m PyInstaller --clean --noconfirm desktop_app.spec