2025 — ongoing

email triage

Agents that read an inbox the way I would, so the things needing a decision surface and the rest files itself.

Private repository — the writing here describes the approach rather than linking to code.

  • Python
  • IMAP
  • LLM agents

Email is a queue nobody designed, with no priority field and no definition of done. The volume is not the problem; the problem is that the few items genuinely needing a decision are indistinguishable at a glance from the many that do not.

This ingests over IMAP and classifies with a mix of deterministic rules and model judgement. The split matters and is the same split that runs through all these tools: rules are cheap, predictable and auditable, so they handle everything they can, and the model is reserved for the cases where intent actually has to be read. Running a language model over every message in a mailbox is slow, expensive and mostly wasted on newsletters.

Classified mail triggers downstream work. It files into hum, and it feeds the document pipelines directly, since a good proportion of what bills and tax consume arrives as an attachment rather than on paper.

It is deliberately read-mostly. The job is to surface what needs a decision and move the attachments where they belong — not to become somewhere I answer mail. It began as imap-tools, a broader toolkit, and the narrow part is what proved worth building first.

The richer client is still ahead of it rather than abandoned, and I may or may not get to it. What is available now does not do what I want while also keeping my mail private, which is the whole reason to consider writing one. It is a much larger job than this, and I am not pretending otherwise by listing it.

The repository is private.