Klorn

Klorn assigns every inbound email exactly one lane — PUSH, MEETING, QUEUE, INFO or SILENT — and most of the time the decision is "you don't need to see this."

The design bet: the LLM never decides. It scores four features (confidence, sender trust, reversibility, urgency), and a deterministic, unit-tested rule maps those scores to a lane — so the decision is auditable without the model in the loop, and every row shows its reason in plain language ("You replied to this sender 6 times", "Payment due today"). Correct a decision and it learns; correct the same sender twice and it becomes an explicit rule.

Irreversible actions — send, delete, forward — require an approval that pins the exact payload bytes with sha256, re-verified at execution. Any drift and it fails closed. That is enforcement in code, not a prompt asking the model to behave.

Self-hosting is a first-class path, not a demo: AGPL-3.0, docker compose up -d with prebuilt images (amd64 and arm64), your own Google OAuth client, your own model key — or point it at a local Ollama, LM Studio or vLLM endpoint and no email content ever leaves your machine. Hard daily LLM spend caps are enforced in code. With no LLM at all, the deterministic keyword floor still classifies at 78%.

Free forever self-hosted. Hosted tier at $8.99/month with a 7-day free trial.