Your Own AI v0.4.1
Private AI on your own machine - offline by default, online only when you choose.
Offline models load on every Windows install
- Fixed: on Windows installs outside the C: drive, the bundled engine
could not start, so offline models never loaded - the app reached
"Starting server with model" and went silent. The engine was located
through a path resolver that fails on custom-drive installs; it is now
found relative to the app itself, the same way the record-keeping
engines already were. Thanks to the tester whose patient diagnostic
reports over three days led us to it. - Graphics-card probes are bounded: a driver that hangs while enumerating
devices no longer stalls every model load - after 15 seconds the load
continues on the processor. - Every way a model load can fail is now named in the diagnostic report,
including a server that starts and produces no output at all.
Deleting a conversation works for every AI
- Fixed: deleting a conversation from an AI's records did nothing for AIs
created after the app first learned to delete - new AIs, AIs restored
from your Vault, and every fresh install. The ability is now given to
each AI's records the moment they are set up, existing ones are brought
up to date on the next launch, and a delete that fails says so instead
of closing quietly.
Your records, warming up honestly
- After launch, the Memory page, the conversations drawer, and the home
screen's "Continue last conversation" all show that your records are
still warming up instead of an empty list, a false zero, or a dead
click - and open the moment they are ready. - Warmup copy says "ready", not "online" - these records never leave your
device.
Attachments
- Attached documents appear as file chips in your message, never as their
extracted text - in regular chats and in project sessions alike. - Attachments stay on your device unless "Send attachments to online
models" is on. The setting now lives in Settings > Routing, because it
decides both where the message goes and what may leave the machine, and
the one setting governs regular chats and project sessions.
Agent sessions, seen clearly
- Background work is named and visible: condensing working notes,
extracting memories, and scripts still running in the background stay in
the rail - even across a stopped and restarted turn. - The rail names the real tool behind an agent's action ("Remember
something for this project") instead of a generic "Use Tool". - Agent sessions start faster: the online model list is cached for a
minute instead of being fetched on every step.
Routing and models
- A paused model is out of automatic routing's reach until you resume it,
and the pause and resume tooltips say exactly that. - Signed-in online accounts stay signed in as the service rotates their
sign-in tokens.
Under the hood
- Rust dependencies audited and updated with the record-keeping crates
left byte-identical; the dependency lockfile is now committed so every
build is reproducible. - Front-end build tooling updated; the development build targets the same
JavaScript level as the production build.
Downloads
See assets below for platform-specific installers. Verify any download against SHA256SUMS.txt.