janitor-template — a daily email-cleanup bot for any inbox

A fully de-personalized, self-hostable copy of a production Gmail "janitor." It runs once a day, triages new mail, files it on a visibility-first schedule (you see important things first; records file instantly), keeps your inbox a clean waiting-room, sends you a daily digest, and learns your preferences over time. Plug in your own account + rules and go.

Nothing is ever permanently deleted (trash only, recoverable). It never touches starred / protected / "nostalgia" mail, and a PreToolUse hook (.claude/hooks/guard.sh) hard-blocks destructive operations even if the model errs.

What it does each run

1. Fetches only new mail since last run (History API cursor). 2. Triages each: junk → trash + one-click unsubscribe + filter (spam only for actual scam); keeper → routes to its label-home. 3. Marks read only disposable noise (receipts/OTP/promo/fluff). 4. Files records instantly; holds important mail visible for a 7-day window, then files + surfaces it. 5. Sends you an email-to-self 📬 Digest (filed / surfacing / needs-you / junk-cleared + trash-chain transparency). 6. Commits its memory so it learns.

How you talk to it

• Reply to any email, slap the #agent-feedback label on it → the bot obeys + remembers next run. #agent-undo reverses its last action on a message.
• Or edit memory/feedback_inbox.md. Or reply DONE on a digest item to dismiss it.

Setup

1. OAuth — create a Google Cloud OAuth client, enable the Gmail API, scopes gmail.modify + gmail.settings.basic; get a refresh token into token.json (gitignored — never commit it). Set OWNER_EMAIL to your address.
2. Your rules — edit memory/rules.md: your label-homes, your protected senders (family/financial/etc.), your nostalgia keeps. Create the labels in Gmail (incl. #agent-feedback, #agent-undo, Digest, and hidden zz-Surfacing/zz-Pinned).
3. Test locally — export OWNER_EMAIL=you@example.com GMAIL_CLIENT_ID=… GMAIL_CLIENT_SECRET=… GMAIL_REFRESH_TOKEN=…; source scripts/lib.sh; get_access_token then follow PROMPT.md.
4. Deploy — private repo + a daily scheduler (a Claude routine, or cron) with the OAuth creds in the prompt/env. One supervised run, then hands-off.

Replace these placeholders

{{OWNER_NAME}} · {{OWNER_EMAIL}} (or the ${OWNER_EMAIL} env var) · {{GITHUB_OWNER}} · {{REPO}} · {{family}} · plus the PROTECTED + NOSTALGIA lists in memory/rules.md.

Files

CLAUDE.md (identity + invariants + the canonical fleet doctrine) · PROMPT.md (the daily procedure) · memory/rules.md (your sender→label map) · scripts/lib.sh (Gmail helpers, no pip deps) · .claude/hooks/guard.sh (hard safety) · .github/workflows/auto-merge-claude.yml.

This is a generic template — no personal data. Adapt the providers/labels to your needs.