The problem
Email volume past ~50 messages/day exceeds what most people can triage well. Hosted services like Superhuman or HEY work but ship your inbox to a third party. Self-hosted lets you keep the data while still gaining the leverage.
Recommended setup
| Agent | Hermes Agent |
| Hardware | Raspberry Pi 5 (8 GB) or Hetzner CX22 (€5/month) |
| LLM | Claude 4.5 Sonnet via Anthropic API (or OpenRouter) |
How it works
Hermes Agent connects to your IMAP/JMAP inbox via a custom MCP tool. The tool fetches headers + first 500 characters of new mail, the agent classifies (action / reply / archive / spam) and either replies (drafts saved to drafts folder, never auto-sent) or files. Approval flow on every send.
Reality check
This is a setup to try, not one we have measured. Expect routine mail (receipts, notifications, scheduling) to be classified better than threads that need context, and keep every send behind your approval. Count the agent's mistakes over the first weeks before you let it file on its own. The API bill grows with the number of messages it reads: estimate it from your daily volume and the model's per-token price before you switch it on.
What breaks
- Long, threaded conversations (agent loses context past ~10 turns)
- Highly contextual replies that need knowledge from outside the email itself
- Anything legal, financial or relationship-sensitive (always reply yourself)
Alternative setups
Nanobot (it lists email among its channels) if you want a lighter personal agent; switch on its shell sandbox and workspace guard first, both off by default. ZeroClaw + a local model such as Mistral 7B if you accept weaker classification in exchange for cloud-free operation.
Agent defaults read in the projects' own documentation on 2026-10-18: Nanobot README (v0.3.5), Nanobot security settings (v0.3.5).