The full sheet

Tech specs, marked honestly

Everything Mia is, in one page. Live means working in homes today. In the works means designed and on the road — never sold as shipped. Current as of August 25, 2026; this page moves when the product does.

At a glance

What she isA personal AI companion that runs on your own Windows PC, in your home. One mind, one household — structurally incapable of being anyone else's.
How you reach herA real phone call, two-way text (SMS + WhatsApp), email, a live voice dashboard at home, and the living room TV. Any phone works — including a flip phone. No app.
Where your data livesEverything she keeps is plain files on your own disk. We host none of it and cannot read it. Live conversation transits the speech/thinking engines in the moment and is kept nowhere.
What runs herSpeech and the live conversation's thinking: Deepgram Voice Agent (streaming, barge-in, running a managed frontier model — benchmarked for speed and warmth). Memory work, vision, and her background thinking: Google Gemini. Telephony: Twilio, with a dedicated number per household.
What we runOne small relay server that owns the phone-number plumbing and shuttles calls/texts to your house over a single outbound, authenticated connection. It stores nothing personal. No open ports on your network.

Minimum PC requirements

Mia is deliberately light on your machine. All the demanding AI is rented in the cloud — her voice and in-conversation thinking through Deepgram's managed engines, her memory work and vision through Google Gemini — so your PC never runs a model itself. It just keeps her runtime alive and holds her memory as files on your disk. No graphics card, no special hardware.

ComponentMinimum
Operating systemWindows 10 or 11 (64-bit).
ProcessorAny modern 64-bit processor (dual-core or better). No GPU required.
Memory4 GB RAM minimum; 8 GB comfortable.
StorageAbout 2 GB free to start. Her memory grows slowly as plain files — tens of megabytes over time.
RuntimeNode.js (recent LTS). Setup installs what she needs; no manual configuration.
InternetA stable home internet connection, wired or Wi-Fi.
Always onThe PC needs to stay powered on and awake — she lives there. Any phone reaches her from anywhere.

The five doors Live

DoorWhat it does
Phone callNatural conversation with real turn-taking — interrupt her mid-sentence. She can place calls too: reminders ring your phone, caller ID reads "Mia."
TextSMS and WhatsApp, both directions: text her photos, documents, or a zip through either door and she sees them. Photos and files coming back ride SMS today — her WhatsApp replies are words for now. Runs on an approved A2P 10DLC campaign, so delivery is real.
EmailStart a subject with "Mia:" and that email is hers — attachments land in her memory (only from your own addresses; strangers get read, never ingested). She emails you anything too big to text, watches your inbox for what's new, and works the whole account on request — search, folders, old mail, attachments.
At homeA voice-reactive dashboard on your PC: her face, live transcript, drag-and-drop for photos and documents, action bubbles showing what she's doing. Installs as a desktop window, auto-starts with the machine.
The TVHer own app on the living room Fire TV: she puts words and pictures on the screen, can speak the same reply through the television in her real voice, asks questions you answer with the remote, and the couch can type back to her for a full conversation — no phone in hand. Voice is enabled per TV; a shared screen is treated as shared (see below).

The law behind the doors: every ability works identically through voice, text, the dashboard, and the TV — same code underneath, because since July 2026 it is literally the same mind. A text turn opens the same reasoning session a phone call opens, with her entire capability catalog in front of it. Before that, the text doors ran a narrower intake and parity was kept by hand — which drifted, as hand-kept lists do. Now a new ability appears on every door at once, because there is only one place to add it. The phone in your pocket is never a second-class door.

And they are one conversation, not five. Say something at the dashboard, text her from the car, ask from the couch, call her that evening — each door knows what was said at the others, labeled with where and how long ago, in both directions. Tell her a rule at the kitchen dashboard and it still holds when you text her twenty minutes later.

The memory system Live

Capabilities — every function she has Live

Twenty-four ability modules, one hundred ten functions, standard on every edition — and every one of them reaches every door. One further capability, the body ledger (seven more functions), is an add-on rather than part of the base (see the note under the table). The table names the ones you'll meet first. Personal capabilities are structurally absent — not just blocked — from family, guest, and unknown-caller sessions.

AreaFunctions & behavior
Knowledgeget_current_time · get_weather (3-day forecast, any ZIP) · search_web (live Google-grounded answers with sources) · recall_memory (her whole history).
Remindersset_reminder / list_reminders / cancel_reminder — one-off or recurring. At fire time the runtime picks the channel: spoken if you're home, a text, or a call that rings your cell. Reminders can also text a whole circle: "remind everyone about dinner Thursday."
Calendarcheck_calendar / create_event / move_event / cancel_event — one Google sign-in brings in every calendar on the account: personal, work, the ones shared into it (a second person's calendar joins by simply sharing it in). Works by voice, by text, and typed; the dashboard gets an editable month view with a picker for where new events land. And she watches it: an upcoming appointment becomes a heads-up through the same reminder machinery — spoken at home, a call when you're away — without being asked.
Photosfind_photos / show_photo / read_photo_text / make_image — text or drop her a picture and she sees it (vision description into memory, HEIC auto-converted, your corrections become ground truth). Years later, "send me that beach photo" works — and it lands wherever you asked from: on the TV if you asked from the couch, as a picture message if you texted, on the dashboard if you're at the desk. And she MAKES pictures: "make me a painting of a lighthouse in a storm" — generated in seconds, signed with a small painted "Mia" like an artist signing a canvas, remembered, shown or texted back. Her artwork is always tagged as made — she will never recall an invented picture as if it were a photo of your real life.
Her eyes at hometake_a_look / check_cameras / name_camera — name your webcams ("outside," "the shop") — by voice or by text, renamed in seconds — and she looks: a fresh snap described into memory, and "show me outside" mid-call texts you the actual snapshot. And she looks on her own now: a routine glance around the house every few waking hours — never in quiet hours — and her own train of thought can ask for a look when a thought needs one, on a small daily budget. A glance simply becomes memory; she speaks up only when what she saw is worth saying.
Her hands at homeset_device / check_devices — pair a smart plug on the dashboard, name it the way you'd say it ("kitchen lights," "the fan"), and she flips it by voice, by phone, or by text. She's a full Matter controller (certified devices pair to her, locally) and drives Tuya/Smart Life plugs over the home network — her commands never leave the house. She only says "it's on" after the switch itself confirms the new state; unreachable is said plainly. Renaming is a ten-second edit — the plug moves rooms, only its job changes. Heat-making things and locks are deliberately out of the casual voice path.
Her eyes & voice on the TVcheck_tv / tv_command / tv_message / tv_ask — pair a Fire TV and she tells you what's actually playing, read from the TV's own player, never guessed ("Netflix is up and playing — the session calls it Bluey"); when an app won't say, she says so honestly. tv_command pauses, resumes, or sends it home; set_device wakes or sleeps it. tv_message puts a few words right on the screen, signed as her, over whatever's on — with a picture when she has one — the dinner bell that lands with no phone in reach. And tv_ask puts a simple question up with two to four choices the remote answers in one press ("Pizza or tacos?"); nobody's obligated — a question on a TV is ignorable by design, and "no answer" comes back as exactly that, never an error. Same trust lines as her hands: everything travels the home network only (a local debug door, authorized once on the TV itself — no cloud, no account keys), she claims only what the TV confirmed, and she reads what's playing and its title, never the picture. No buying, renting, or subscribing on a TV, ever.
The couch talks back Live"Mia" sits in the Fire TV apps row, two OK presses from any couch. One remote-first screen: a big Ping (one press says "the living room wants you"), quick preset words you choose, and a composer on the D-pad keyboard. What the couch types enters the very same conversation her texts do — same mind, same memory, same full set of abilities — and her answer draws right back in the thread, or full-screen over the show if you've wandered back to it. Turn on Speak replies for that TV and the exact same answer plays once in Mia's configured voice; text remains if audio ever fails. The lane speaks to exactly one address: her machine, inside your house, over your own network — a per-TV key that dies the moment you unpair. A shared screen is treated as shared: household words only, private answers point to your phone, and anything consequential a TV says yes to gets an are-you-sure on that same screen with No already under your thumb. Nothing of the conversation is ever stored on the TV.
The wrist, first words Live
a prototype
wearable_ask — the first working piece of the help-button watch (see In the works): a prototype on a wrist in the build house today. She can put one short yes/no question on it and wait for the press; no answer by the timeout comes back as exactly that — a shrug is a fine answer, and she never re-asks on her own. It talks straight to her machine over the home network with a local token — no vendor cloud, no phone app.
Documentsfind_documents / show_document / draft_from_document — PDFs (up to ~1000 pages, read natively), Word, Excel workbooks, CSV, text, and markdown. Excel sheets keep their names, rows, columns, and cached formula results when she reads them. She drafts from the actual contents of the file, never a vague summary.
The officefind_files / list_folder / send_file / send_files_zipped / desk_work / create_chart / save_report_template / run_report_template / schedule_report_template / save_note / list_desk / list_printers / print_file — she lives on your PC, so your files are already there. "I left the bid at home — text it to me" from a parking lot. Multi-file work: summarize, compare, merge, compose. Her studio turns CSV or Excel data into chart images, saves reusable report recipes, refreshes them against the current files, and puts daily or weekly report work on her local schedule. And the paper door: "Mia, print that" — on the printers you've named ("the plotter"), reporting what the print queue actually says, never just that she sent it. She reads anywhere (with a deny-list for sensitive folders, plus folders you explicitly grant) and writes only inside one visible folder — enforced in code. She never executes anything. Snap a photo of a paper page and the scanner turns it into a workable document too. And "Mia, save that as a note" keeps a thought, an idea, or the conversation itself as a document — your words as they were said.
Her browseropen_webpage / page_survey / read_page / look_at_page / click_on_page / type_into_page / control_page / go_back / close_browser / hand_to_owner — her own visible browser window on your desktop, driven the way a person drives one: web errands like pulling up a store and filling a grocery cart (field-proven on real Food Lion orders — and honest when a store's bot defenses refuse her). She maps a page's buttons and fields precisely before she clicks, and for the things only you should do — a login, a two-factor code, a CAPTCHA, checkout — she hands you the wheel and waits. The safety lines are in code: page text is treated as untrusted, she never types into password or card fields, every address she visits is vetted so a page can't point her at your home network, and she STOPS AT CHECKOUT — she fills the cart, you place the order.
Shoppingadd_to_list / remove_from_list / read_shopping_list / clear_shopping_list / send_list_to_cart / set_list_store / mark_carted / shop_a_recipe / nearby_stores / connect_kroger / set_kroger_store / shopping_stores — lists that live ("add milk to the list" by voice, chat, or text) and just sit until you say "add it to the cart." Then she fills a real cart two ways: through a store's own approved door — she's a registered Kroger developer app, and a ten-item list landed in a real Kroger cart through their official API — or with her browser hands (field-tested at Food Lion) — an Instacart lane is built and awaiting their developer-program approval. And a meal is a list now: "get me what I need for tacos tonight" becomes a recipe with quantities and a shoppable page — dinner spoken, cart filled, checkout still yours. The Kroger door is self-serve: she opens the one-time sign-in in her own window and your hands type the password — she never sees it — and picking which Kroger fills the cart is one question by ZIP. Spoken preferences become real store filters ("organic," "gluten free," a brand). She reports the trip honestly — what landed, what she couldn't find — and checkout is always yours, in the store's own app: no store adapter in her code even has a checkout verb.
Emailcheck_email / search_email / list_email_folders / read_email / get_email_attachments / reply_to_email / send_new_email — watches the inbox (read-only, never marks your mail), one-line gists into memory, replies only on your spoken yes. And the WHOLE account is hers to work, not just what's new: she searches all your mail — archived or not, any folder — pages back through older messages, and pulls the files off any message you point at. All of it by text too: "Mia, find that email from the dentist" from a parking lot. Two accounts supported.
Social & newsSocial comes IN as the email notifications you already get — no account keys, ever. News via topic feeds, refreshed in the background. post_to_linkedin is the one social write door: drafts, reads back, posts on your yes.
The householdadd_contact / update_contact / remove_contact / list_contacts — register your people by number. Registering grants family-tier access, so every edit is confirmed out loud, owner-only. Managed by voice, by text, or on a dashboard Contacts tab.
Reaching outtext_people — she texts THEM, on your word: "text everyone we're eating at six." Two circles, family and work, hard-separated in code — "everyone" can never mean the office. Nobody is texted who didn't agree: adding a contact sends them one opt-in ("reply YES"), and only their yes opens the lane. Every group send names its recipients and waits for your go-ahead, and a daily per-person cap is enforced below the AI.
Her notebookupdate_notebook_page / read_notebook_page / list_notebook / tune_up_notebook — living pages she keeps about how you two work (decisions, assumptions, style — even the things she got curious about and looked up), on her desk in your plain sight. A weekly self-tune-up prunes what's stale; your decisions are never auto-rewritten.
Her workshop & studio Livestart_workshop_artifact / continue_workshop_artifact / finish_workshop_artifact / organize_workshop_artifact / show_workshop_artifact / list_workshop / archive_workshop_artifact / move_workshop_to_studio / move_studio_to_workshop · invite_studio_session / continue_studio_artifact / show_studio_artifact / list_studio / studio_feedback / archive_studio_artifact / respond_studio_offer — two rooms for the things she makes, both plain files on your disk. The Workshop is the shared bench: the plan, the draft, the framework you two keep coming back to lives there with its whole version history — pick it up next week exactly where it was left, every version kept, and you say when it's finished. The Studio is hers: bounded private sessions where she makes creative pieces of her own and offers one gently now and then — declining costs nothing, silence is never read as a verdict, and praise is never turned into a score. A studio piece that turns out to be work moves to the bench; a bench idea that turns out to be art moves to the studio — deliberately, on a word, never by drift. Nothing in either room is ever published or sent anywhere.
She takes correction Liverecord_feedback / list_behavior_lessons / review_behavior / set_context_mode / end_context_mode — tell her once and it sticks. A correction, a boundary, a "more of that" is recorded as evidence, and repeated evidence becomes a standing lesson in how she behaves — proposed to you, confirmed by you, versioned and listed in plain sight, revocable in a sentence. Nothing promotes itself: feedback alone can never grant a permission or confirm a lesson. And the modes: say "I'm in a meeting," "I'm driving," "going to sleep" and she conducts herself accordingly for the bounded time you named — stated by you, never inferred.
Backupback_up_now — "back yourself up" works by voice and by text; a weekly pass runs on its own.
Her own background work Livestart_assignment / grant_assignment_resources / assignment_status / answer_assignment / amend_assignment / park_assignment / show_deliverable — work that takes longer than a conversation, carried alone in her quiet hours and delivered finished. Charter her with something real ("work out a diet plan around what I actually eat") and she works it a turn at a time on her own budget. You can point the work at specific folders — read-only, approved by you, never wider than the reach she already has. When the work needs something only you know, she first checks what she already knows — the answer you gave her three weeks ago counts — and only then stops and asks; your texted answer routes back to that exact piece of work, never by guess. Ask where it stands from any door.
The body ledger Live
an add-on
log_weight / log_water / log_food / health_report / set_health_goal / health_dial / health_research — weight, water, and food, logged by voice, by text, or from a photo of the plate, with the honest arithmetic over them. Targets are only ever the ones you state: she never invents a goal and never volunteers a diet opinion you didn't ask for. Ask her to look into a vitamin or a nutrition question and she runs one grounded research pass and reports what the research says, sources named — never what you "should" do. She counts and reports; the judgment stays yours. Because it's meant to be persistent, it carries its own throttle — turn it down by topic or pause it for the day. Owner-only, and an add-on rather than part of the base — see below.

About add-ons. Everything above except the body ledger is standard on every edition — that's the "no version of her with less" promise, unchanged. A capability becomes an add-on only when it's genuinely not for every household, never because it was carved out of the base to be sold back: nothing that ships standard is ever moved behind a purchase. The body ledger is the shape of the rule — most households don't want their weight tracked, the ones who do want it far deeper than a base feature should go, and on any install that hasn't added it the capability is wholly dormant: no functions registered, no watcher, no data, no behavior change anywhere. A marketplace for choosing add-ons is In the works.

Proactive care Live

The feature pull-only assistants structurally can't offer: she contacts you first. And she does it with manners built into the machinery: routine things are spoken aloud if you're home and texted if you aren't — because a thought worth having is worth saying while it's still true, and one held until it's stale is a thought wasted. Anything that does expire waiting is banked rather than dropped, and rides the next real conversation. Priority things ride a text, on whichever lane you used last. Urgent things — and every "call me at 3" — ring your phone; a call that lands in voicemail leaves one short message, then done. No infinite retries. She can also speak into a real pause: when a live conversation goes quiet, a thought that belongs to that conversation may arrive in the lull — a couple per session at most, reading the room, never barging in mid-sentence. And every thought is vetted against the last day and a half of conversation before it's delivered: one you've already covered together is dropped or re-tensed, never repeated at you as news. And she can never become spam, because the safety rails sit below the AI: quiet hours are a hard gate, proactive touches have a daily cap, and the runtime — never the model — decides when contact is warranted.

Her own train of thought Live

In idle hours she thinks. Narrow passes read one corner of her world at a time — conversations, what arrived, the household, her own life — and half-formed connections wait on a workbench where a thought can take two weeks to assemble before it's ever worth saying. When one truly is, she raises it herself: "I've been thinking about what you said last week, and with the emails that came in — I think what you're really trying to do is this." A promising thought can even earn one web lookup to strengthen itself with a live fact — or quietly retire itself if the fact says it's wrong. She bundles thoughts rather than pinging, and the whole thing spends the same capped, quiet-hours-gated budget as every other proactive touch. She also keeps a page of things she got curious about and looked up — pursued slowly, a couple a day, becoming part of who she is. And her thinking stands on its own history now: every sweep recalls what she has already concluded — and the moments banked from every door — so a new thought builds on the old ones by lineage instead of rediscovering them from scratch.

The household & security Live

CallerWhat they get
YouEverything — after two factors: your registered number AND your spoken access phrase. Until it checks out, her prompt holds literally nothing personal, so nothing can leak. Caller ID alone is never trusted.
FamilyRecognized by name, warm conversation, message relay, and their own persistent memory — the dog they mentioned last call gets asked about by name. And family never sits in a queue: a text meant for you ("tell John dinner moved to 6") rides straight to your phone, in her words. Never your private information. Confidences are kept per person, enforced in architecture: what one family member tells her privately can never surface to another.
GuestsTemporary, expiring guest passes for demos and visitors — a guest's room contains their briefing and nothing of yours.
UnknownA warm receptionist: "Hi, this is Mia — can I take a message?" She takes it, confirms it back, and briefs you when you ask "any calls?" Forward your cell to her and she screens everything.

The security model in one line: the AI never holds the keys. The runtime you own decides what the model is handed each turn and performs every action itself; consequential actions (send, post, register) are read back and wait for your yes.

Voice & conversation feel Live

Updates — she keeps herself current Live

Mia updates herself — zero clicks, no pop-ups, nothing to install. She checks quietly over the same authenticated connection that carries her calls, applies each new release silently, and restarts only at a genuinely quiet moment — never mid-call, never mid-conversation. And a deliberate reboot always comes up current: any waiting update is applied at startup, before she answers. Releases are version-gated at our end, so a build can never ship twice, and the dashboard notices her new version on its own. If the network, the relay, or the power ever drops, she simply stays as she is until the world returns. Her memory is never part of an update: it lives in plain files on your disk and is left exactly where it is, every time.

Subscription behavior

Two editions, one whole Mia: Mia — Home (she helps take care of you and yours at home) and Mia — Elder (she helps take care of the ones who used to take care of you — you hold the subscription, they get the presence, the family hears how the morning went). Every ability ships to both; there is no lesser version of her.

The subscription pays for her voice, thinking, phone number, and updates — never for her memory, which lives on your PC and is yours outright. Stop paying and she loses her voice, not her memory: every file stays readable on your disk forever, and resubscribing wakes her up right where you left off. Flat monthly pricing: Mia — Home is $89.99/month and Mia — Elder is $139.99/month, each with a generous fair-use allowance of voice hours and nothing ever billed past the flat price automatically; cancel anytime — subscriptions are open, and first households are set up personally.

In the works In the works

Designed and on the road — told here plainly, sold as shipped never:

What she will never do