msgsift

Local-first email triage that turns your inbox into a calendar-style todo list.

msgsift fetches your unread mail over IMAP, classifies each message with an LLM (local Ollama by default, or Cloudflare Workers AI), and organizes everything into a per-day todo list with a small web UI. A nightly job writes you a short recap of the day and rolls unfinished items forward. Your own manual todos live in the same list. Nothing leaves your machine unless you opt into the Cloudflare backend.

How it works

• Classify every unread email as ACTION_REQUIRED, FYI, NEWSLETTER, or IGNORE, with a one-line reason and (where useful) a suggested reply.
• VIP list — senders (by email or domain) are always forced to ACTION_REQUIRED.
• Forwarding hints — content rules suggest who to forward something to.
• Daily rollover — at end of day, FYIs/newsletters auto-close, unfinished action items carry to tomorrow, and finished ones stay (scratched out) as a record of what you did.
• Nightly recap — the model writes a couple of sentences summarizing the day.

Install

uv tool install msgsift

This gives you three commands: msgsift (fetch + classify), msgsift-nightly (recap + rollover), and msgsift-web (the web UI).

Configure

Config lives in ~/.config/msgsift/ (override with MSGSIFT_CONFIG_DIR):

mkdir -p ~/.config/msgsift
cp config.example.toml ~/.config/msgsift/config.toml
cp credentials.example.toml ~/.config/msgsift/credentials.toml
# then edit both — config.toml for settings, credentials.toml for secrets

credentials.toml holds only secrets (IMAP passwords, API tokens) and should never be committed or shared.

Run

msgsift          # fetch unread mail, classify, store, print a digest
msgsift-web      # serve the web UI (default http://127.0.0.1:8000)
msgsift-nightly  # write the day's recap and roll items over

Typically msgsift runs on a timer through the day, msgsift-nightly once at night, and msgsift-web as a long-running service. Example systemd user units are in systemd/.

Backends

• Ollama (default): fully local and private. Point [classifier.ollama] at your Ollama server and pull a small model (e.g. qwen2.5:3b).
• Cloudflare Workers AI: offloads inference to a larger model. Set [classifier] backend = "cloudflare" and provide an account_id + api_token. Note: with this backend, message envelopes (sender, subject, snippet) leave your device.

License

AGPL-3.0-or-later.