A persistent, interruptible AI assistant that lives on the desktop, on the phone, and in the cloud — with memory it can be taught, and hands it can be trusted with.
1. What it is
XELA is an assistant with four parts that work as one:
Layer
What it is
Brain
A Cloudflare Worker with 45 stateful components (Durable Objects) driving the cognitive pipeline — comprehension, memory, emotion weighting, planning, and a language stage. Stateful and conversation-aware, and interruptible mid-thought.
Model layer
A companion worker that routes to Workers AI models for language, vision, speech-to-text and text-to-speech. No third-party API keys to manage.
Desktop app
A floating, always-on-top mascot on Windows. Click to talk, drag to move, tray icon, wake word, screen and camera awareness.
Mobile app
An installable web app (PWA) on any modern phone. Chat, voice, camera, tasks and memory, no store approval needed.
It is not a chat window with a skin. The assistant has a persistent self: what it remembers, what it is allowed to touch, how it speaks, and which brain it talks to are all things the owner controls.
2. What it can do to itself
Self-management, all of it visible to the owner:
Remember durable facts — when something worth keeping comes up, it writes a memory. Every memory can be edited or deleted by the owner.
Assign itself follow-up workopt-in — with autonomy enabled, it can create a task for itself, then complete it later and report back in the conversation.
Notice and speak firstopt-in — it samples the screen periodically, and only speaks when something meaningfully changed. Quiet during focused work by design.
Answer to its name — saying “Xela, Xela” brings it forward from anywhere.
Persist its own configuration — capabilities, reply style, tone, and which brain endpoint it uses survive restarts.
Change its voice — reply length and tone are adjustable, and take effect on the next message.
What it cannot do to itself: it cannot rewrite its own instructions or model weights, cannot deploy or replicate itself, and cannot grant itself a permission. Every capability toggle belongs to the owner, not to the assistant.
3. What it can do to others
Actions that reach outside the app, each behind an explicit permission and an audit entry:
Capability
Scope
Guard
Open links, apps, files
Anything the OS can launch
Permission + confirmation per action
Type and send keys
The currently focused window
Permission + confirmation, keystroke log
Control smart devices
HTTP switches (Shelly, Tasmota, ESPHome), Home Assistant, camera snapshots
Dedicated IoT permission, per-action audit
Write files
Its own sandbox folder only
Path confinement; cannot write to system paths
Notify
OS notifications for due tasks
None needed; read-only
It never contacts third parties on its own. There is no email sending, no messaging, no posting. Anything that leaves the device is either a model request to the operator's own cloud worker, or an action the user approved in that moment.
4. What it does on the user's behalf
Conversation with context — every reply is informed by open tasks, assigned roles, remembered facts, and current interaction mode.
Reads entire answers aloud — long replies are spoken in full, not truncated.
Hands-free conversation — a live voice loop listens in short windows, transcribes, answers, and speaks back.
Voice access — “open…”, “search…”, “type…”, “press…” executed on the desktop after confirmation.
Sees — camera frames and screen captures are described, including what app is open and what needs attention.
Writes — generates and saves documents and code into its workspace on request.
Remembers and reminds — tasks with due dates produce one-time notifications.
Texts in bursts — long answers arrive as a few short messages, the way a person replies.
Takes direction mid-answer — you can add to a request while it is still working; it either folds the addition in, or stops and asks which you meant.
Follows the roomopt-in — it transcribes what it overhears, recognises enrolled voices, and decides for itself whether to speak, stay quiet, or answer a specific person. It can decline out loud with a single word.
5. Where it differs from general assistants
XELA
Browser / OS assistants
Presence
Floats above other windows, tray, wake word, always available
Inside a browser tab or a fixed panel
Memory
Persistent, inspectable, editable by the owner
Mostly per-conversation or opaque
Interruption
You can change your mind mid-generation; it adapts or asks
Typically queue the message until the current reply ends
Hands
Files, apps, keystrokes, smart devices — under permission
Rarely beyond the browser sandbox
Ownership
Your Cloudflare account, your data, your model choices, switchable brain endpoint
Vendor-hosted, one size
Cost shape
Usage-based infrastructure, no per-seat licence
Per-seat subscription
6. Deployment and data
Hosting: Cloudflare Workers, Durable Objects, R2, D1 and KV — no servers to patch.
Desktop data: tasks, memories and settings in a local JSON file under the user's profile.
Mobile data: tasks and memories in the device's local storage.
Model traffic: text, audio and images are sent to the operator's own Cloudflare worker for inference. Nothing is sent to a third-party AI vendor directly.
Distribution: desktop installer via direct download; mobile via a link, installable to the home screen.
7. Honest limitations
Anything offered to a customer should come with these stated up front:
Not multi-tenant yet. One brain instance per deployment. There is no billing, no user accounts, no per-customer isolation.
No HIPAA / BAA posture. It should not hold patient health information in its current form.
No practice-management integration. It does not yet book into Open Dental, Dentrix or similar — this is the single biggest gap for clinical use.
Wake word needs a microphone and is unreliable when a phone screen is locked; continuous listening is opt-in and off by default.
Voice recognition is not biometric — it separates two or three recurring voices on one microphone, and can confuse two people with similar pitch. It must never gate access to anything.
Room listening is foreground-only in a browser — a phone stops capturing when the screen locks; the desktop app keeps listening while it runs.
IoT coverage is three protocols. No Tuya/SmartLife cloud, no smartwatches, no PTZ camera control.
No calendar sync. Scheduling is local; calendar integration needs OAuth per provider.
Screen awareness is periodic, not continuous, and deliberately disabled during focused work.
It cannot click inside other applications. Voice access opens, types and sends keys — but UI automation of arbitrary apps is not implemented.
8. Who it fits today
Best suited to a single professional or a small practice that wants one assistant across desk and phone, with memory of their work and the ability to act on it — and an owner who is comfortable granting permissions deliberately.
Not yet suited to: multi-location enterprises needing central administration, regulated clinical data handling, or teams expecting per-seat licensing and SSO.