YouOS — Personal Email Copilot

YouOS is a full local Python app (not an instruction-only snippet). It drafts email replies in your style, grounded in your real past replies.

Naming: YouOS is the shared app/package. During setup it personalizes its name to OS (e.g. BaherOS) for your local instance at YOUOS_DATA_DIR=~/YouOS-Instances// — that's a local config/data directory, not a fork of the project.

Safety & impact

Read these before installing:

• Sensitive data ingested locally. YouOS reads your Gmail (sent + threads), Google Docs, and optional WhatsApp exports into a local SQLite DB. This stays on your machine, but it is sensitive content — protect the data directory (YOUOS_DATA_DIR/var/youos.db) accordingly.
• Installation runs local package code. ./scripts/install.sh builds a .venv and installs this repository — review the source (especially scripts/install.sh, pyproject.toml, app/) before running.
• Background persistence is opt-in. YouOS only runs as a launchd LaunchAgent if you explicitly run youos service install. Foreground youos serve does not persist.
• Scheduled/nightly runs are opt-in. Ingestion, fine-tuning, and autoresearch only run automatically if you've enabled the nightly pipeline; youos improve is the manual equivalent.
• External model fallback is optional. Drafting is on-device by default. Cloud fallback only fires if review.draft_model or model.fallback is set to a cloud model — set model.fallback: none and review.draft_model: local for strict local-only operation.

Install and runtime model

• Install is manual: run ./scripts/install.sh (Python 3.11+) — it creates a .venv, installs YouOS, sets up the on-device model (MLX) on Apple Silicon, and runs the doctor
• Note: this executes local package install code from this repository; review source before installing
• Requires python3 (3.11+) and a Google ingestion backend — gog (default), gws, or native (see Credentials)
• Drafts run on your fine-tuned local model by default (Qwen + your LoRA, served warm); the cloud is only a cold-start before your model is trained, or a fallback
• Optional runtime path override: YOUOS_DATA_DIR

Credentials and configuration

• Required: a Google ingestion backend for Gmail/Docs — gog (default), gws, or native (youos[google] + OAuth); set ingestion.google_backend
• Optional: Claude CLI/API credentials only for the cold-start/fallback (review.draft_model / model.fallback)
• For strict local-only: set review.draft_model: local and model.fallback: none

Trigger phrases

• "draft a reply to this email"
• "write this email for me"
• "how would I respond to this"
• "what would I say to"
• "help me reply"
• "draft in my style"
• "youos"
• "my email copilot"
• "email copilot"
• "my copilot"
• "generate a draft"
• "reply draft" / "email draft" / "draft reply"
• "compose reply"
• "write a response"
• "email response"
• "how do I usually reply to"
• "reply to this"
• "help me write"
• "write an email"
• "compose a response"
• "email assistant"
• "my writing style"
• "train on my emails"

Requirements

• Apple Silicon Mac (M1/M2/M3/M4) with 8GB+ RAM (16GB recommended) — the local model runs on MLX (./scripts/install.sh sets it up)
• Python 3.11+
• A Google ingestion backend for Gmail/Docs: gog CLI (default), Google's gws CLI, or the native Google API (youos[google] + OAuth)
• ~5GB free disk space
• Run the UI locally by default (do not expose publicly unless intentionally secured)

Quick start

# Install (creates .venv, installs YouOS + MLX on Apple Silicon, runs the doctor)
cd youos
./scripts/install.sh
source .venv/bin/activate

# Check system requirements (Python, Google backend, MLX, disk space, etc.)
youos doctor

# Run setup wizard (identity, ingestion, style analysis) — or open /welcome in the browser
youos setup

# Draft a reply (uses your local fine-tuned model by default)
youos draft "paste inbound email here"
youos draft --sender john@company.com "email text"

# Run the web UI (then open /feedback, /stats, /settings, /about)
youos serve
youos service install     # or run it as a background service (starts at login)

# Compare the backends on YOUR mail, ranked by how closely each sounds like you
youos compare-models --limit 30 --semantic

# Warm local-model server (loaded once for fast drafting)
youos model server status

# Check status / view stats
youos status
youos stats

# Run the nightly pipeline manually (add --verbose for step-by-step output)
youos improve --verbose

# Golden benchmark (10 curated cases)
youos eval --golden

# Full corpus health report (pairs, quality scores, top senders)
youos corpus

# Ingest a WhatsApp chat export (optional — augments your corpus)
youos ingest --whatsapp ~/Downloads/WhatsApp-Chat.txt

# Add a sender note (immediately rebuilds their profile)
youos note john@company.com "integration partner, prefers bullet points"

# Submit a feedback pair directly from the terminal
youos feedback --inbound "email text" --reply "your reply" --rating 4

# Teardown (remove all data, keep code)
youos teardown

Drafting inside Gmail

Install the YouOS browser extension (Chrome/Edge/Brave) — it lives in the repo's

extension/ folder (homepage),

and the web UI's Gmail page has one-click "Load unpacked" steps. The extension adds a

panel to Gmail:

• Open an email → click the teal ✉ launcher → the panel opens
• Sender + message auto-detected; add an instruction or pick a tone
• Click Generate → drafted in your voice; Insert into Gmail drops it in the reply box
• Rate 1–5 and Submit feedback — YouOS learns from it, same as the Review Queue

A bookmarklet remains as a no-install fallback (it can break when Gmail changes its markup;

the extension doesn't).

How it works

1. Ingests Gmail, Google Docs, WhatsApp exports — plus organic pairs from emails you sent without YouOS
2. Builds a retrieval index — BM25 + query expansion + semantic (LRU-cached) + multi-intent + per-account isolation + same-thread 2× + subject + topic signals + sender-type boosts + quality scores + relative confidence thresholds
3. When you ask for a draft: detects multi-intent, retrieves score-ranked thread-deduplicated exemplars (reply preserved 600 chars, inbound trimmed 400), prompt token budget enforced; generates with per-mode persona + first-name greeting on your fine-tuned local model by default (Qwen + your LoRA, served warm via mlx_lm.server so it's fast and on-device). The cloud is only a cold-start (before your model is trained) or a fallback; a per-draft model badge + the Stats "Drafting with" row always show which model actually ran
4. Every email you review trains the model further — curriculum-ordered, quality-filtered, training pairs deduplicated by similarity, DPO pairs supported; nightly pipeline skips steps when data insufficient
5. Nightly: ingests + organic pairs, incremental persona re-analysis (90-day weighted, EWMA avg words, p25/p75 confidence intervals), fine-tunes (with golden eval check), runs autoresearch on rotating benchmark sample
6. Autoresearch benchmarks rotate weekly (seeded re-sample) — prevents overfitting to fixed test cases; golden eval composite tracked in pipeline log
7. Style drift detection: Stats dashboard flags when your writing patterns shift significantly
8. Your best-rated, least-edited replies surface higher in future retrievals via quality scoring
9. Sender profiles track reply-time patterns and topics; youos note immediately rebuilds that contact's profile
10. Submit feedback from terminal: youos feedback --inbound "..." --reply "..." --rating 4
11. Setup wizard asks for internal domains — accurate sender classification from day one
12. Facts store (/api/facts) — save context about contacts, projects, and preferences; facts are injected into generation prompts automatically for context-aware drafts
13. Auto fact extraction — sender notes and feedback notes are parsed automatically on save using 15+ rule patterns (preferences, timezone, schedule, sign-offs, roles, relationships, project metadata); negation-aware with confidence scoring; LLM (Claude CLI) fallback for unstructured notes; fact deduplication/merging on upsert
14. Measure it, don't guess — youos compare-models drafts your held-out replies under each backend and scores them against what you actually wrote (voice-match), so you can verify the local fine-tuned model beats the cloud on your mail
15. Readiness gate — a soft "preparing your voice model" banner holds you back from relying on drafts until your model is trained and benchmarked (with a "Run benchmark now" action); drafting still works meanwhile
16. Becomes your OS — during setup the app personalizes its name from your first name (e.g. Baher → BaherOS)

Security & privacy notes

• Gmail/Docs ingestion uses your configured Google backend's auth (gog / gws / native OAuth); review connected accounts before ingestion
• Drafting is on-device by default; the cloud is only the cold-start/fallback. If it's used (review.draft_model: claude or model.fallback: claude), inbound email/context is sent to Claude for that draft
• For strict local-only operation, set review.draft_model: local and model.fallback: none in youos_config.yaml
• Data location defaults to local instance paths under YOUOS_DATA_DIR (e.g. ~/YouOS-Instances//), or the repo's var/
• Review PRIVACY.md before first ingestion/deployment

Provenance

• Source/homepage: https://github.com/DrBaher/youos
• This skill bundles a full local Python app and is intended for explicit local install/review before use.