Garvis

A local, edge-only personal triage + chief-of-staff assistant. Garvis sweeps your Gmail, Outlook, text messages (Google Messages), and optionally WhatsApp through local MCP servers, triages everything, cleans out the noise, hands you a prioritized briefing — and can talk to you out loud.

Everything runs on your machine: a local Ollama model does the judgment, local MCP servers do the I/O, and (optionally) a local faster-whisper + say voice daemon lets you ask "Garvis, what's open?" Nothing is sent to a third-party LLM.

What it does each run

1. Gather new mail/texts (and WhatsApp, if enabled) since the last run.
2. Classify every item — PROMOTION / UPDATE / ACTIONABLE / PERSONAL / WAITING / CONCLUDED / UNSURE — governed by your config/rules.md.
3. Clean up — soft-delete promotions/updates/concluded email (recoverable from Trash ~30 days), guarded by deterministic protection rules so important mail is never touched. WhatsApp cleanup is separate and off by default: only PROMOTION chats, only when allow_whatsapp_delete: true, and it clears a whole conversation (delete-for-me — see Safety). SMS is classified but never auto-deleted.
4. Prioritize what's left into a ranked, chief-of-staff briefing.
5. Deliver a dated digest to digests/ and email a copy to you.

It remembers across runs (SQLite): a recoverable deletion audit, sticky loops (a settled thread can't be silently reopened by a bad run), and a small memory graph of the people and context in your life that sharpens triage.

Design principle

The LLM never drives tool calls. Python orchestrates MCP calls deterministically; the local model only classifies, summarizes, and prioritizes. Deterministic safety guards (VIP senders, protected keywords, OTP grace windows) live in code, never in the model.

Quick start

cp config.example.yaml config.yaml                 # then edit with your values
cp config/profile.example.yaml config/profile.yaml # optional: the memory graph
python3 -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt
ollama pull qwen2.5:14b                             # local judgment model

python -m garvis.run --check                        # connectivity check
python -m garvis.run                                # full sweep (DRY-RUN by default)

You also need the local MCP servers it talks to — personal-gmail, personal-outlook, and google-messages, plus optional whatsapp-mcp — cloned and built, with their paths set in config.yaml (WhatsApp is off by default; enable it under mcp_servers).

See STANDALONE.md for the architecture, full setup, the voice daemon, and the durable-state / memory-graph design.

Safety

• Dry-run by default — classifies and writes the digest but deletes nothing until you set dry_run: false.
• Email is soft-delete only — Gmail/Outlook removals go to Trash, recoverable for ~30 days.
• WhatsApp is the exception — a WhatsApp cleanup is delete-for-me of an entire conversation and is not recoverable like Trash (it never affects the other person and is never delete-for-everyone). It needs a second opt-in beyond dry_run (allow_whatsapp_delete: true), applies only to PROMOTION chats, and is off by default.
• Protected items never touched, and every deletion is logged with sender, subject, reason, and id.

License

MIT — see LICENSE.