Howland privacy policy

Revision 2

Howland is self-hosted software that runs on hardware you own. Ayers Electronics Inc does not operate a cloud service for it, there are no Howland accounts, and we collect nothing about you, your files, or your questions.

What we collect

Nothing. Howland contains no analytics, no telemetry, no advertising or tracking software, and no crash reporting that phones home. There is no Howland account and no login to any service we operate, because no such service exists. Your conversations, your files, your searches, and your settings stay on the machines you run Howland on.

Where your data lives

Howland is built so that nothing has to leave your network. The models run on your own server. The library is served from your own drives. The coding agent reads and edits files on the machine you run it on, and the thinking happens on your server.

Chats, artifacts, projects, and settings are stored locally on those machines. Deleting Howland deletes its data.

What Howland remembers

Howland keeps a memory of what you tell it, so you do not have to repeat yourself. It is worth knowing exactly how that works, because memory is the most sensitive thing the product does.

It lives on your server. Remembered facts are plain text files on your own Howland server, with a search index built from them. Nothing about them leaves your network. Conversations are no longer only in the moment: what matters from them is kept.

It fills itself. A helper on each computer you work from reads that machine's own sessions and sends new turns to your server, which decides what is worth keeping. Only what you typed is considered, never the assistant's replies, and code, diffs, shell output, and stack traces are thrown away.

Secrets are stripped before anything is written. Anything shaped like an API key, a password, a private key, a token, or a web address with credentials in it is removed before the text is stored or indexed.

You can have memory nobody else can read. A household member can keep a private scope, sealed with a key derived from their passphrase. Both the note and its search index are encrypted, and other members, including whoever administers the machine, cannot read it. The limits are worth stating plainly: while a private scope is unlocked its contents are in the server's working memory, so anyone with administrator access to that machine could read them; backups contain the wrapped key, so a weak passphrase could be attacked offline; and if the passphrase and the recovery code are both lost, that memory is gone, including to us.

Search runs on your hardware. Memory search uses a small embedding model downloaded once at install from its public source and run unchanged on your server's processor.

You can read it, change it, or delete it. Everything remembered is listed on the server's memory page, where any item can be edited or removed, and memory can be switched off entirely, for one member, or for a single session.

What the software connects to

The client apps connect to one thing you choose: your own Howland server, at the address you enter. Beyond your own server, Howland contacts only what you configure or ask for:

• Your own server's services. The model brain, the offline library, the map, and the dashboards are all served by your own server on your own network.
• Web search, when you use it. Your server runs its own search service, installed with it rather than reaching out to anyone else's. The agent's ability to search is off until you turn it on; when a search does run, that service queries public search engines from your server, and Howland attaches nothing about who you are.
• Model and library downloads, at install. Setup fetches the models and the archives you choose from their public sources, or installs them from a drive you prepared offline. After that, no connection is required for normal use.
• GitHub, for release checks and downloads. The server tray checks the public release page once a day for a newer version, and the desktop app checks the same page for its own updates — both checks are what the auto-update preference controls, and turning it off stops them; you can still check for updates yourself at any time. Both are plain requests that send nothing about you, and nothing installs without your say-so.
• Cloud model pass-through, only if you enable it. If you enable cloud model pass-through and only then, prompts and files you send to a cloud model are transmitted to the provider you chose (such as Anthropic or OpenAI) under that provider's terms; Howland counts the token usage those providers report, locally, to enforce the spending limits you set, and nothing else leaves your machine.

No training on your data

Your conversations and files are never sent to us and are never used to train any model. The models Howland runs are open-weight models published by third parties, downloaded once and run unchanged on your hardware.

The same is true of everything Howland remembers. Remembered facts stay on your server, are never sent anywhere, and are never used to train any model.

Children's privacy

Howland does not collect personal information from anyone, including children.

Third parties

We do not sell, share, or transfer any information to third parties, because we collect none. There are no third-party data processors. The optional connections listed above go to public sources or to services running on your own hardware.

Changes

If this policy changes, the updated version is posted at this address under a new revision number. The address does not change.