MiCasa

The starting point deserves credit, the product deserves its own identity, and sovereignty needs a clear meaning in everyday life.

MiCasa is the app. Foundation-LifeStyle does the continuing work.

“Agentic” means that the backend can carry a job forward rather than only reply to a message. Foundation-LifeStyle keeps track of what the household asked for, what information may be used, what still needs doing, and which decisions belong to a person. It can put Scouts to work and bring useful changes back to MiCasa after the original conversation has ended.

Take a simple request: “Watch this warranty and remind me before it runs out.” MiCasa is where a person shares the document, chooses whether the job is personal or shared, and sees the reminder. Foundation-LifeStyle uses the permitted document, keeps track of the deadline, gives the Scout its job and limits, and remembers any correction. If the next step is booking a repair or spending money, that needs its own permission and a supported service connection.

We customized the Buzz frontend to fit Foundation-LifeStyle: personal and household spaces, sharing, service management, and the views people need to follow their twins’ work. Foundation-LifeStyle supplies the work behind those screens. Its memory, Scouts, service connections, privacy protections, and upkeep do not all live inside the fork. Together, the customized frontend and agentic backend make the complete household experience.

What sovereignty means at home

MiCasa treats household context as something the household stewards. Personal conversations, shared family material, and the home itself belong to different audiences, and the experience should make those boundaries understandable.

Sovereignty does not require a cloud-free life. It means the household decides when an outside service is useful, what it may receive, and what it may do. A twin can draw on outside intelligence without opening the family archive or a resident’s private memory to that service.

The aim is practical control: private and shared spaces that make sense, actions that respect the person or household authorizing them, and choices that can be changed as family life changes.

Cloud help, with household privacy

We call this a semantic air gap. In everyday terms, the twin prepares a limited question for an outside AI, removes details it does not need, and replaces identifying information with neutral labels. It brings the answer home, where the twin can reconnect it with the private details that make it useful. Foundation already includes the gateway that does this work.

Imagine asking why a bedroom is taking longer to cool. Outside AI might help compare possible causes from a carefully prepared description. It does not need the family’s names, address, calendar, wearable readings, or private conversations. The twin keeps those details at home and uses the returned answer alongside the household’s own records.

For MiCasa, the rule is simple: if the privacy check finds a problem, or cannot be completed, the question does not go to the cloud. The twin can try to help using what is available at home, ask for a narrower question, or explain why it cannot continue. Getting an answer never justifies quietly giving up the household’s privacy.

Where MiCasa runs

MiCasa is designed to be home-hosted on a Mac mini or supported Android hardware. Android is a host path in its own right, not a phone client dependent on a Mac-hosted system. The household can choose the form that fits its home.

The set-top box opportunity

Home set-top boxes are central to our approach to wider access through telecommunications providers. That familiar, always-on device offers a practical home for sovereign AI, with an open-weight small language model providing local intelligence close to the household and its connected services. Availability on that hardware depends on the provider and supported home setup.

AI at home is the household’s first source of help. When a more capable cloud service is useful, the same privacy rule applies: prepare a limited question, keep the household’s private history and memory at home, and send nothing if the check cannot pass.

What we inherited

Buzz is Block’s open-source app and agent-native network foundation. It provides the upstream basis for people and agents to participate in a shared communication environment. MiCasa inherits that foundation. We did not invent Buzz, and our work does not transfer ownership of Block’s code to us.

MiCasa preserves the applicable Apache License 2.0 and Block copyright notices. Calling it a fork makes the relationship clear: Buzz is the open-source foundation, while MiCasa is the household product we are shaping from it. Block’s authorship of Buzz does not imply endorsement of MiCasa.

What we changed in MiCasa

MiCasa turns that foundation into a household experience: choosing a personal or shared context, bringing material into a conversation, managing Apps & Services, and seeing useful views for things such as weather, hazards, media, and the home itself.

Across platforms, the experience stays native to the device while preserving the same household identity, context, and sense of control.

How Foundation-LifeStyle completes the product

MiCasa makes the choices and results visible. Foundation-LifeStyle connects those choices to memory, continuing work, the home’s devices, and the services the household has allowed. A helpful conversation becomes useful over time because the backend can remember the purpose, respect the limits, and follow through.

Follow one household request from MiCasa into the continuing work of Foundation-LifeStyle.

MiCasa is where a person chooses a context, brings in material, talks to a twin, and sees the result in a form that fits the task.

Foundation-LifeStyle, the home substrate, carries the twin’s work and memory beyond one conversation, connecting the household to the services it chooses. AI running at home can use the household’s private context. Outside AI receives only a question prepared and checked for that purpose.

Scouts are continuing work with a purpose, scope, and lifecycle. They can keep a bounded job moving after the person closes the app, return when something changes, and ask for help when a decision belongs to a person.

PACT: ask without opening the room

PACT is how one twin asks another twin for one clearly defined thing without gaining access to that twin’s private context. The request states who is asking, what is needed, what may be returned, and when the request expires. The receiving twin can answer, refuse, or ask its person to decide.

For example, one twin can ask, “Is there a shared two-hour opening on Friday evening?” The other twin can return an available time without exposing the calendar entries, meeting titles, locations, or reasons behind every blocked hour. PACT carries the narrow question and the permitted answer, not a doorway into the other person’s room.

HIVE: reuse a finding without exposing why it was found

HIVE is the household’s way to share an observation that another authorized Scout can safely reuse. The shared item includes the finding, where it came from, how fresh it is, and when it should expire. It does not include the private conversation that caused a Scout to look, the resident’s hidden goal, or source material that belongs to a personal room.

If one Scout confirms that an appliance model has been recalled, another household Scout should not have to repeat the search. It can reuse the sourced recall finding while learning nothing about which resident first asked, why they asked, or what else sits in that resident’s context. HIVE reduces duplicate work without turning shared learning into shared surveillance.

LogOS: household oversight without an operations console

LogOS is the plain-language view of continuing work. It tells the household which Scouts are running, paused, finished, or waiting for a person; which source is unavailable; what limit or approval is holding an action; what changed; and whether anything remains running after a job should have ended.

It does not reveal the private content a Scout used to do its work. A household administrator can see that a personal Scout is healthy or needs attention without seeing the resident’s messages, wearable readings, documents, or findings. The aim is oversight without surveillance, expressed as decisions a person can understand rather than logs they must decipher.

A million agents can take a request and return a result. That is table stakes. MiCasa’s moat is privacy: Scouts can do useful work close to the rich context of a home without flattening the home, its residents, and their private lives into one giant account.

In plain language, a Scout is a persistent, purpose-bound AI worker you give a job, not a conversation you have to keep open. It can wake when relevant information changes, use only the context admitted for that job, keep state between runs, work inside agreed authority and resource limits, bring back meaningful changes, ask when a decision belongs to a person, and retire when the intended outcome is complete.

Residents can have personal Scouts that work for them alone. The home can have household Scouts that look after shared concerns. They may coordinate, but one does not inherit the other’s private context. For the wider strategy and why Scouts, rather than one-off agents, become the unit of continuing and compounding work, read Beyond the Agent: Why Scouts Scale Agentic AI.

A Scout belongs to somebody

A Scout is an agentic worker with a named purpose, an owner, an audience, an admitted set of sources, a limit on what it may do, and a condition for stopping. A personal Scout works for one resident. A household Scout works for the home. The distinction determines what the Scout may know, what it may change, and who may see the result.

This is not one omniscient household bot. A Scout watching a resident’s wearable can use signals that person consented to share without exposing them to the household. A Scout looking after the heat pump can know the building without opening anyone’s private conversation. MiCasa can pass a permitted constraint between those worlds, such as a preferred bedroom temperature, without revealing the private history behind it.

The moat is private household context

The value comes from working across the domains that make a home a home while preserving the boundary around each one: smart-home systems, personal wearables, household documents, photographs and video, settlement and delivery, hazards relevant to the home or a consenting person’s current area, and learning that adapts to an individual.

A Scout sees only what its job and audience permit. Asking cloud AI for help does not expand that permission. The Scout’s question must pass the same privacy check, with unnecessary personal details removed and household memory kept at home. If it cannot pass, the Scout must find a permitted way forward or explain the limit.

The chat can end; the Scout keeps its place

A resident can say, “Keep watching this warranty until the replacement arrives,” “Tell us if this typhoon changes what the household should do,” or “Help me master this before the exam.” MiCasa turns that intent into a Scout with a visible purpose, scope, resource limit, audience, and stopping condition.

The Scout moves into the home substrate. It keeps enough state to know what it has checked, what changed, what remains unknown, and which decision still belongs to a person. The phone can lock and the conversation can leave the screen; the work still has a governed place to live.

Quiet, steerable, and able to finish

A Scout should not become a daily stream of “still looking” messages. It can wake when a relevant source or household condition changes, compare the new state with what it already knows, and return only when something material happened or a person needs to decide. MiCasa can still show that it is alive, when it last checked, and which source is stale.

“Pause until Monday.” “Only use official sources.” “Do not spend anything.” “Include the house, but keep my health context private.” A resident can steer the Scout in ordinary language without rebuilding it as a rule set. When the intended result arrives, the Scout returns what changed, what remains uncertain, and the source or receipt behind the outcome. It then releases its temporary authority and closes, leaving nothing quietly running.

That is the distinction: not another agent waiting for the next prompt, but private agentic workers that carry the intent of a household and its residents across time without taking ownership away from the people who live there.

Personal and household contexts have different audiences

A person can have a private relationship with their twin while also participating in a shared household context. MiCasa is designed to keep those experiences distinct without making family collaboration feel technical.

A family might share a calendar, an appliance manual, or a plan for a trip without sharing every private conversation or the personal reasons behind each preference.

Foundation-LifeStyle keeps personal memory separately from the shared household record. The shared space is built from what people choose to contribute, rather than by pooling everyone’s private life and hiding parts of it on screen. Looking after the home does not, by itself, give someone access to another resident’s private relationship with their twin.

A connection comes with choices

Connecting a calendar should not mean handing over the whole diary. A twin may need to answer “Are you free on Friday?” without sharing meeting titles, locations, or the reasons someone is busy. Reading a document, showing an answer to the family, and sending a copy outside the home are separate decisions.

Trust also depends on what is actually connected. Foundation-LifeStyle is designed to check that the app and service connections are the approved versions, as well as checking what the person has allowed them to do. A familiar name or icon is not enough. When access changes, the next use must respect the new permission, not an old yes.

Authority belongs to the action

A twin can suggest useful work without automatically receiving permission to perform every step. The action itself should carry the household’s chosen limits, approvals, and expiry.

Unknown remains an outcome

A failed sensor reading is not a temperature of zero. A disconnected calendar is not an empty calendar. MiCasa should make uncertainty visible so a person does not have to infer it from a spinner or an optimistic sentence.

Results should remain understandable

A household should be able to tell whether something was suggested, approved, attempted, completed, or still waiting. The language of the app should make those distinctions feel natural.

Without memory, a twin is a stranger at the start of every conversation. With one undifferentiated memory, it becomes a privacy problem. MiCasa gives memory a structure: what happened, what should persist, whose context it belongs to, how the relationship may adapt, and what must be corrected or forgotten.

Durable memory carries what should survive the chat

A personal twin can remember facts, preferences, commitments, people, places, devices, routines, and corrections that belong to its person. At the beginning of a new conversation, it retrieves only the small set relevant to the request rather than pouring an entire life into the model’s context.

That memory helps the twin continue a commitment, honor a preference, recognize the device someone means, or avoid repeating a corrected mistake. It remains context, never authority. Remembering that someone prefers the bedroom cooler does not grant permission to change the thermostat, spend money, disclose the preference, or act for another resident.

Unfinished commitments belong in memory too. A reminder that has not been delivered, a decision still waiting for approval, and a promise to follow up should survive the move to a new conversation. Shortening an old chat must not silently cancel the job or turn “waiting for your answer” into permission to proceed.

Athena learns from outcomes

Athena is the outcome-learning layer. It asks not only what the twin did, but whether the result was useful. A completed task, a failed intervention, an ignored suggestion, a correction, or explicit feedback can become evidence about what worked for that person and situation.

Athena does not let the twin grade itself and call the lesson true. The outcome has to be established independently or supplied through trusted feedback. Athena can then carry the lesson forward: notify less often, prefer a better source, recognize that a particular kind of suggestion was mistimed, or repeat a pattern that produced a verified result.

Character memory lets the relationship grow

Character memory governs how a twin expresses itself. It begins with a stable character and can make bounded, consented adjustments in warmth, directness, playfulness, curiosity, caution, initiative, formality, detail, and gentleness. Athena and signed human feedback provide evidence; character memory decides whether that evidence is strong and appropriate enough to change the twin’s interaction style.

The adaptation can be relationship-specific. The same twin may learn that one resident wants the answer first, another appreciates a gentler explanation, and a shared household room needs a more neutral voice. Those differences do not become a blended household personality, and private feedback from one relationship is not exposed to another.

Character learning cannot change who the twin is, expand what it may access, approve an action, alter a spending limit, or turn affection into authority. A model impression cannot rewrite the character directly. A proposed change is bounded, checked, recorded, and reversible.

Correction and forgetting are part of remembering

A useful memory system needs an ordinary way to say, “that is wrong,” “that changed,” “remember this instead,” or “do not use that again.” A correction supersedes the older memory rather than competing with it forever. Expiring information can age out. A person can withdraw consent for character learning, and the twin’s future behavior should be replayed without the withdrawn influence.

Suppose the twin keeps describing a room as a nursery after it has become a study. Correcting the memory should change future answers and suggestions, not just produce an apology in the current chat. Withdrawing a memory means it must no longer guide the twin. Erasing old conversations, backups, or copies already shared is a separate question that the household should be able to understand.

Memory follows its audience. A resident’s twin cannot quietly search another resident’s private memory. Household memory contains what was deliberately placed in the shared context. A Scout receives only the task-local memory its mission needs. Sharing a conclusion does not transfer the private history that produced it.

Learning about you is not training on you

Foundation-LifeStyle separates a twin learning how to help its person from training a general-purpose AI on that person’s life. Its model-training rules exclude household conversations, documents, photographs, and other personal material, even after identifying details have been removed.

A twin can remember that you prefer a short answer without turning that preference into training material for a wider model. Any use of an outside AI still has to respect the household’s choices about what may be sent and how it may be used.

An appliance has more to tell a household than whether it is on.

A useful view can bring together an equipment identifier, available telemetry, the correct manual, a purchase record, and a service history. These are separate sources that have to be joined carefully. A manual is relevant because it names the equipment, not because its filename contains a plausible brand.

MiCasa brings connected devices, equipment records, and home services into one household view. Equipment that is offline or awaiting setup remains visible, so the inventory describes the real house rather than only what answered most recently.

Suppose someone asks why a room takes longer to cool. A credible answer needs comparable observations and enough context to explain the comparison. It should report missing history rather than fabricate a baseline. A suggestion to book service is a proposal; the booking would require its own integration, authority, and confirmation.

Imagine a unit taking twenty-three minutes to cool a bedroom where comparable conditions previously required fourteen. In this example, the service proposal includes the observations: the model from its rating label, the relevant bulletin, the warranty from its receipt, and a suitable calendar slot. The person reviews the proposal; completing a booking requires a supported service connection and confirmation. The numbers illustrate the scenario, not a measured result from a particular household.

Health and wearables are among the most sensitive possible household inputs. A shared home does not imply permission to share a person’s symptoms, medication, or inferred habits.

The personal-signal design centers on the owner’s own history: a sustained change in an approved signal, enough comparable observations to explain it, and a result shown to that person first. Collection and disclosure each need explicit rules. A missing signal stays missing, and a hypothesis about a trend remains distinct from a diagnosis.

A person can choose to share a practical constraint, such as a temperature preference, without sharing the history behind it. A prescription-pickup scenario uses the same principle: the shared calendar receives only the scheduling information it needs. The privacy check considers what those constraints reveal in combination, as well as the details directly disclosed.

A long-term, personal baseline helps someone notice a change and decide what deserves attention. Each health-related use depends on supported data sources, consent, validation, and appropriate professional advice. A personal observation is not a medical assessment.

A household’s most useful material is often unglamorous: the warranty, the invoice, the lease, or the instructions for an appliance. MiCasa helps people ask ordinary questions and return to the document when they need the detail.

Scans, photographs, PDFs, and supported word-processing files belong in the same household conversation. If something is unreadable or still being processed, MiCasa says so instead of filling the gap with confidence.

Coverage is visible in human terms: what is ready, what needs attention, and what is still waiting. Tax packets, bank passbooks, utility bills, and receipts become useful without turning the home into a document-management project.

Removing material should be just as understandable as adding it, with choices that respect household intent and the realities of backups and previously shared copies.

A sovereign household substrate needs its own historian. The archive is not a pile of files waiting to be searched. It is the family’s record of people, places, seasons, celebrations, ordinary days, changing homes, and the stories that connect them.

The Household Historian

MiCasa’s Household Historian gives that record continuity. It brings together photographs, video, speech, family-supplied captions, calendars, household documents, important events, and the changing state of the home into named, bounded episodes. Instead of requiring a person to remember a filename or an exact date, it can work with the way families actually remember: “the week my mother visited,” “the summer the bedroom air conditioner kept failing,” or “her last birthday at the old house.”

The historian does more than retrieve an item. It returns an answer with the original material close at hand, keeps the people and audience attached to the episode, and lets the household correct a name, relationship, date, place, or interpretation. A correction should become part of the family record so the same mistake does not have to be fixed in every future retelling.

The household remains the editor of its own history. MiCasa can connect what the archive supports, distinguish a family-supplied story from a machine observation, and leave uncertainty visible when the connection is not yet known. That discipline makes the history more useful, not less alive.

History the household can revise

A date printed on a photograph, a relative’s recollection, and a twin’s guess about the occasion are different kinds of information. The historian should keep that difference clear, so a repeated guess does not gradually become a family fact. If a date is unknown, the memory can stay undated until someone helps fill it in.

The history also has to respect changing permission. If a photograph or document is no longer available to the shared household, an old summary should not become a way around that choice. The historian should revisit the story, remove what it can no longer support, and make the gap clear. A family can correct its history without giving up control of the material behind it.

The archive should come alive

The historian turns long-term memory into recurring household experiences rather than waiting silently for a search box.

Time Machine extends that history into the room. When the moment is right, MiCasa can propose a bounded set of photographs for a shared screen and explain why they belong together, while the people present remain in control of whether anything is shown.

MiCasa’s share and result surfaces make photos, audio, and video part of the household conversation.

A useful question might be “what did the builder say about the timeline?” MiCasa can help find the moment in the recording and return the clip, keeping the original experience close at hand.

Indexing metadata, recognizing content, identifying a person, and releasing a copy outside the household are different operations. Consent and authority should be evaluated for the operation actually being attempted. Approval to index an album does not by itself authorize every inference or onward disclosure.

Processing and sharing are explicit choices. Household media is handled at home wherever the selected feature supports it. Asking cloud AI for help is not permission to upload the family album or a private recording: the twin must prepare a limited question that can pass the privacy check without sending those originals. Deliberately sharing a photograph is a separate choice, with control over who receives it and whether location and camera details are removed.

These media capabilities give the Household Historian material it can return to naturally: the original photograph, the exact moment in a recording, the words on a page, or the family episode that connects them. Search finds the material; the historian helps it remain part of family life.

In Japan, a household intelligence system must be ready for the ground, water, weather, and mountains to change the day without warning.

The author and the team behind MiCasa live in Japan. Earthquakes, tsunami, landslides, flash floods, volcanic eruptions, typhoons, extreme heat, and severe storms are not remote categories here. They can affect one family member at home, another at work, and another travelling through a different municipality at the same time. The hazard module is therefore not a convenience feature. It is one of the clearest reasons for MiCasa to exist.

Two locations, one household picture

The home location anchors the building and the people who may be there. A consenting person’s current mobile location anchors the conditions around them while they are away. MiCasa can use both without collapsing them into a continuously shared family tracking map.

When a warning arrives, the useful questions are immediate and human: Does it affect the house? Who may be in the area? Has everyone seen it? Does someone need a check-in? Which official instruction applies where each person actually is?

From warning to coordinated action

Home and Current Location views can bring forecasts, earthquake information, evacuation notices, and place-specific warnings into the household context. A hazard card can make the affected place, urgency, source, and next step clear. It can invite acknowledgement, help the family check in, and keep the relevant official guidance one touch away.

A tap cannot resolve an emergency, and MiCasa should never pretend otherwise. Official instructions remain authoritative. The role of the household twin is to make those instructions easier to understand in context, help family members coordinate, and keep uncertainty visible rather than inventing reassurance.

A quiet screen must have a clear meaning. “No current warning from the sources checked” is different from “the last update is old,” “the source is unavailable,” or “this kind of warning is not covered here.” The household needs to know which is true, especially when deciding whether to check official guidance directly.

Share the warning, keep the family’s whereabouts private

A public typhoon bulletin can be collected once and reused instead of every personal twin repeatedly fetching the same notice. Each household can then work out privately whether it affects the home or a consenting resident. Sharing that public work does not require sharing the family’s movements or the reason one person is concerned.

Safety without turning location into surveillance

Location is sensitive household data. MiCasa should evaluate hazards close to the household boundary whenever possible, retain only what the household has chosen, and reveal a person’s whereabouts only to the audience and degree needed for the safety purpose. A family can coordinate during danger without making continuous movement history everyone’s default view.

This is also where intelligence at home matters. AI on a supported home host can relate an official warning to the people and places the household cares about. The set-top-box design follows the same principle. Outside AI may help explain public guidance, but that does not give it the family’s location history or private memory. If a question cannot pass the privacy check, it stays at home; the family can still turn directly to the official guidance already available.

A household trip is a useful design test. People can share dates, budget limits, and practical preferences without supplying every private reason behind them. A twin could help assemble options from that permitted information and show where each option came from.

The travel design connects planning to a commitment the household controls. In this scenario, a family asks for somewhere warm during a school break, within a budget and with a kitchen. Each twin contributes permitted constraints, and the household twin adds what the home needs: a service visit already arranged, for example. The result brings those preferences together into clear options.

A personal Scout can watch an approved ticket source within a written limit and expiry, then bring back an opportunity when it appears. Booking a household trip needs broader approval: the agreed itinerary, spending authority, provider terms, and the adults required to approve it.

A candidate itinerary, an approved proposal, a submitted booking, and a confirmed trip are different states. Completing a purchase depends on an available booking connection and the household’s permission. A plan is not a confirmed reservation.

The payment design gives each owner a separately governed wallet and the household a shared wallet with its own approval rules. Allowances start at zero. Permission covers a specific amount, recipient, purpose, and expiry; the keys that authorize payment stay separate from the AI that reasons about it.

A payment stays within that permission. Raising a shared spending limit requires the household’s chosen approval rather than whichever agent happens to ask. Actual payment depends on a supported provider and an enabled, approved payment connection.

Agent-to-agent payment can extend that model to small, bounded services. The household should always be able to distinguish a proposal, an approved payment, and a completed exchange.

Ambiguous outcomes deserve care. A lost connection after payment must not trigger a second charge just because the first confirmation did not arrive. The twin needs to check what happened or leave the result clearly unresolved. Paying and receiving are also different events: a payment receipt does not, by itself, prove that the promised service was delivered.

The learning-companion design uses a goal, approved material, and a private record to help someone return to a difficult concept over time. Progress means what the learner can recall and use later, as well as which lessons they have opened.

In this design, a learning Scout works around a stated goal and visible progress. It revisits material, adjusts the pace when the learner struggles, and keeps the learning material close at hand.

The adult cases matter alongside the children’s: a professional certification around a working week, a language before a move, or learning enough about a subject to ask better questions of a specialist. The job has a completion point and retires with a useful record.

Specialist tutoring remains a development area, with adaptive scheduling, consent, and appropriate handling of children’s data as essential requirements.

A person may want more than an answer now: watch an approved source, follow a developing subject, or keep a bounded collection current. MiCasa lets that request become continuing work without turning it into an unlimited mandate.

The experience begins with ordinary language. A person describes the outcome, sees who owns the job, how long it may run, and what it may do, then pauses, changes, or retires it through the conversation.

A job may become blocked, expire, pause, or complete. MiCasa should keep those states distinct so an old record is not mistaken for work that is still moving.

The same caution applies to work suggested by a twin. A useful idea does not become a commitment until the owner agrees to its purpose and limits.

“Watch for a part for this appliance” is an example of continuing work with limits. Researching a subject, following a planning application, watching a source for change, or maintaining a family archive uses the same discipline. Where a job also needs a purchase or another outside action, its description identifies the additional permission and service connection it needs.

When a job needs several lines of work, the substrate can coordinate them under one outcome and budget. The person should still receive a manageable result rather than a flood of internal activity.

A household should not need an IT department. The people who live there should not have to understand services, queues, storage pressure, expired connections, model health, or software releases simply to keep their home intelligence working.

Bee is the autonomous healer of the Household Agentic Substrate. It watches whether the substrate is doing useful work, not merely whether a process claims to be running. It looks for stale connections, blocked work, missing results, resource pressure, failed updates, and the quieter failure in which everything appears green while the household receives nothing.

Understand the problem before trying a repair

Bee looks for what stopped working, what else might be affected, and which repairs the household has already allowed. It can then choose a limited repair or explain why a person needs to decide. When the cause is uncertain, it should investigate or ask for help rather than guess.

Routine recovery can happen without asking the household the same question every time. A change that needs new access, could destroy something, or affects a physical device still needs the appropriate permission. Bee’s job is to look after Foundation-LifeStyle while the household stays in control.

Repairs within agreed limits

The household can give Bee standing permission for a narrow set of upkeep tasks: reconnecting a failed service, clearing temporary working space, or restoring a version that worked before. The limits say what Bee may touch, how long it may try, and what must remain untouched.

A successful repair does not earn wider access. Fixing a photo connection, for example, is not permission to read the family album or change who can see it.

A repair is not complete until something else verifies it

Bee does not mark its own work healthy. After a repair, an independent check must confirm that the intended household capability is useful again. If the check fails, Bee can roll back, try another already-authorized recovery path, or stop and bring the evidence to a person. “Restarted” is not the same as “healed.”

That check also gives Bee something worth learning. A connection that lights up green but still brings in no new material has not recovered. A repair that causes another problem should not become the preferred answer next time. Bee’s learning should follow what actually worked, so the household benefits from experience without granting the healer more power.

Healing without opening private rooms

Bee can see operational health without reading the content that belongs to a resident or household room. It can know that a media index stopped advancing without seeing the photographs, that a wearable connection is stale without reading the measurements, or that a Scout is blocked without reading the private request that created it.

LogOS translates this into household language: what Bee noticed, what it repaired autonomously, what changed afterward, what remains unresolved, and where a person needs to decide. The result is a home system that can care for itself without turning its residents into administrators or turning maintenance into surveillance.

MiCasa grows in understandable stages. Availability depends on the household’s host, connected services, and permissions. A described use is not automatically an enabled service in every home.

Hazard awareness belongs near the foundation of that journey, not at the end of a feature roadmap. The ability to relate an official warning to the household’s home and consenting members’ current locations, then help the family acknowledge and coordinate, is a must-have in Japan and a natural part of MiCasa’s wider purpose.

Working with telecommunications providers can bring that experience to the set-top box already at the center of many homes. An open-weight small language model on that home device can make sovereign AI broadly available without turning the household itself into a cloud account.

Cloud AI has an important role. Foundation’s existing privacy gateway gives MiCasa a way to ask for outside help while keeping the household in control. Every such question has to pass the privacy check; an unavailable or unsuccessful check means it is not sent.

Travel, payments, learning, personal signals, richer media, and broader device coverage belong to the same household design rather than a collection of unrelated features. Each has its own service connections, permissions, and readiness requirements.

The constant across that growth is sovereignty: the household remains the center of context, decisions, and continuing work.

Closing MiCasa should not erase the purpose of a continuing job. A Scout can carry that purpose forward and return when something changes or a decision is needed.

An interruption does not erase the job

A power cut or lost connection can interrupt the work without erasing its purpose. Foundation-LifeStyle keeps the job’s agreed limits, the last confirmed step, and the decisions still waiting for a person. When the service returns, the Scout can work out where to resume rather than ask the household to explain everything again.

Resuming does not mean blindly repeating the last action. If a connection drops just after a request was sent, the Scout must check whether it was accepted before trying again. If that cannot be established, “I do not yet know what happened” is the useful answer. Recovery should preserve both the job and the household’s control.

Help should not multiply the bill

A Scout may need other workers to help it finish. MiCasa’s rule is that the work still belongs to the same person or household and carries the original limits. Dividing a job, retrying a step, or continuing it elsewhere must not create a fresh spending allowance.

The cost of finding an answer and permission to buy something are separate choices. A Scout allowed to research a repair has not been allowed to book or pay for it. When it needs more money, more time, or a different permission, it should make that decision visible to the person who owns the job.

Some jobs can only observe and report. Others can propose or perform an approved action. If a step fails, the Scout can retry when safe, use an allowed alternative, ask for help, or stop. Keeping a mission alive does not require taking an unsafe next step.

Learning from one person’s request does not automatically become household-wide knowledge. MiCasa preserves the audience of the information even when the work itself continues.

Material from a page, document, or service informs the job; it does not quietly expand what the job is allowed to do.

Continuing work also has limits. A job can expire, be revoked, run out of resources, ask for help, or stop. Persistence should serve the household rather than become permanence without consent.

Privacy should not live in a distant settings page. It should be present when a person chooses a context, connects a service, shares material, or approves an action.

The household should be able to understand what is personal, what is shared, what may leave the home, and what can be changed later.

Uncertainty is part of that honesty. MiCasa should say when it needs another observation or a person’s judgment instead of smoothing every gap into a confident answer.

A household-facing agent product has to connect several kinds of work. Someone brings in a piece of information. They choose who it belongs to, connect a service, ask a twin for help, approve an action where required, and understand the result. The experience should feel continuous from one step to the next.

MiCasa is the customized Block Buzz frontend, adapted to fit Foundation-LifeStyle. Foundation-LifeStyle is the agentic backend substrate that gives it memory, Scouts, connected services, and continuing work. The app and the work behind it are distinct parts of the same product.

The modifications are central, not incidental. They are how an open-source communication foundation becomes a household product with its own identity and purpose.

“End-to-end” names the whole journey: from asking in MiCasa, through the work carried out by Foundation-LifeStyle, to a result the household can understand and act on. Both parts develop together.

The lasting value is in the details a household should be able to rely on: private spaces that stay separate, memories that can be corrected, shared warnings that do not expose private movements, and jobs that can survive an interruption without repeating an uncertain action. Helpful learning should not turn family life into training material, and asking more workers to help should not quietly multiply the bill.

In Japan, that promise has a safety dimension. MiCasa should be able to connect earthquakes, tsunami, landslides, floods, volcanic activity, typhoons, and other hazards to the places the family actually cares about, help people check in, and keep official guidance close. A sovereign household system should know enough to help without turning family location into a surveillance product.

Home set-top boxes are the route we see to wider access through telecommunications providers: familiar home hardware with an open-weight small language model close to the people it serves. The aim is sovereign AI as home infrastructure rather than another cloud destination.

Intelligence at home can still benefit from cloud AI. Foundation already provides the privacy gateway, and MiCasa’s direction is clear: ask only the question needed, keep private household memory at home, and stop the outside request if its privacy cannot be checked. The household gains access to more capable help while staying in charge of its own life.

A household deserves both a useful interface and an honest account of the work behind it. That is the standard MiCasa is being built to meet: sovereign AI whose context, authority, outside calls, and claimed outcomes remain visible enough to govern.