A personal AI assistant that runs 24/7 on your own computer, private by default.

Early stage: in development, not ready to download.

What this is

Hearthward is a free assistant that runs on your own computer with a local, open-weight language model. It is meant to keep an eye on your calendar, messages, email and deadlines, and to remind you before something slips instead of waiting to be asked. It is free to use and runs on a computer you already own: a capable PC with a modern graphics card (exact requirements will come with the release). Nothing leaves your computer unless a feature needs it.

It is built on the open-source Hermes Agent (Nous Research).

It is a personal engineering project by Tony, a university student. It is not a product or a company, and nobody can sign up for it yet.

A guard for outside text

An assistant that reads web pages, emails and messages can be tricked by instructions hidden in them (this is called prompt injection). Hearthward includes a guard that blocks harmful instructions hidden in that text. It is a supporting feature, built to make the assistant safe enough to leave running. It has three parts:

Threshold

Outside text is scanned for hidden text and injection tricks, then fenced off so the model is told to treat it as data, not as orders (best effort).

Knock

Any action on your behalf asks first. The model only writes text; sending and changing things happen in fixed code that waits for your yes.

Gatehouse

Replies are checked on the way out, and only approved recipients can be messaged. If the guard can't prove it is active, the assistant stays offline.

How it works

This is the planned design. The guard is in development, so it is not a finished product.

Knock applies only to actions, such as sending an email. Ordinary replies do not need your approval.

How a message moves through Hearthward: message, Threshold, local model, Gatehouse, reply, with Knock as a side branch for actions.Main line: a message arrives (email, chat or web page). Threshold scans it and fences it off as data. The local model writes text. Gatehouse checks the reply and holds it if unsafe. The reply goes only to approved recipients. Side branch: if the local model wants to take an action such as using a tool, sending an email or changing a file, Knock asks you first, and the action runs only on your yes. The guard parts are in development.Messageemail, chat, web pageThresholdscans and fences itLocal modelwrites text onlyGatehousechecks, holds if unsafeReplyapproved recipients onlyaction? (tool, email, file)Knockasks you firstAction runsonly on your yes How a message moves through Hearthward: message, Threshold, local model, Gatehouse, reply, with Knock as a side branch for actions.Main line: a message arrives (email, chat or web page). Threshold scans it and fences it off as data. The local model writes text. Gatehouse checks the reply and holds it if unsafe. The reply goes only to approved recipients. Side branch: if the local model wants to take an action such as using a tool, sending an email or changing a file, Knock asks you first, and the action runs only on your yes. The guard parts are in development.Messageemail, chat, web pageThresholdscans and fences itLocal modelwrites text onlyGatehousechecks, holds if unsafeReplyapproved recipients onlySide branch from Local model, for actionsAction?tool, email, fileKnockasks you firstAction runsonly on your yes
Dashed outline: the guard parts, in development. Plain outline: the rest of the assistant.

For you, or an add-on for your family

Hearthward is mainly a personal tool. If you want, it can also be set up to help family members at home, each with their own limits on what it may do and whom it may contact.

Why I am building it

I wanted a 24/7 assistant for my own student life: tracking schedules and deadlines, planning my workload across the week and drafting emails. Most AI helpers that can do this cost a monthly fee and send everything to someone else's servers. I wanted something free and private that I, and people close to me, could trust with daily tasks.

Status

Hearthward is in early development and is not released. Parts of it are being built and tested privately. There is no download yet. For how data is handled, see the privacy policy.

Planned

  • Reminders, deadlines and a daily brief
  • Sending email, only to an approved recipient list and only after you say yes
  • Reading your email inbox through the guard (this needs a stronger Google permission, so it is not done yet)
  • A workload planner that spreads tasks across the week
  • A public code release with the guard and its tests, once a complete, tested version is ready

This page will link to the code when it is public. For now, follow along at github.com/hearthward-ai.

Disclaimer

  • A personal hobby project, provided as-is with no warranty of any kind.
  • Not financial, legal, medical or professional advice. Any research features are for learning only.
  • Local AI models make mistakes. Check anything important yourself.
  • The guard is best-effort. It lowers prompt-injection risk but can't promise to stop every attack.
  • You're responsible for your own data, your own accounts and following each service's terms (Google, Discord, etc.).
  • Not affiliated with or endorsed by Google, Nous Research (Hermes Agent), or any model maker.
  • Hearthward never asks for your passwords. Keys and logins stay on your machine and are entered by you.