// tomko.ai / Haven
Haven
We want to rethink personal computing around the home: software that understands the information and devices in a household, works naturally across them, and keeps people in control of their own data.
Haven is our local AI system for the home. Haven 1 was our first working implementation: a physical device, an iOS application and a television experience that brought family photographs, search and voice interaction together.
Overview
We started with family photographs.
We wanted a household to have its own place for its memories, with software that made those memories easier to find and enjoy across the devices already in the home.
Haven 1 brought that idea to life. Photographs from an iPhone could be stored on a device at home. A person could search for photographs through Mila, see matching results on their phone and display them on the television. Voice interaction provided another way to navigate the experience.
We built and used Haven 1 in our household and deployed a device to another household in February 2026. The experience worked, but we also learned that making a system easy to set up was only part of the challenge. Making its intelligence useful enough for people to return to it was the harder problem.
Haven 1
Capabilities
Haven 1 ran on a Raspberry Pi 5, with its own local services and an iOS application. During initial setup, the device could create a temporary Wi-Fi network so a household could connect and configure it through the app.
The iOS application transferred photographs, Live Photos and videos to the device over the local network. Previously uploaded items could be skipped, and the files were stored at home.
Mila used an indexed photograph collection to respond to search requests. The system included face detection, location-related information and chronology, although the search experience remained imperfect.
A television connected to Haven 1 could display photographs returned by Mila. Searches made through the iOS application could show results on both the phone and the television.
Haven 1 recognized the wake words Haven, Mila, Alfred and Roger. Spoken requests could trigger photograph searches, and switching agents could change what appeared on the television.
The iOS application included interfaces for Mila, Alfred and Roger. Haven 1 brought these components into one household experience, even though their capabilities were at different stages of development.
How it works
Haven 1 connected four parts of the household experience.
A device in the home
A Raspberry Pi ran the local services, stored household information and connected to the television. Initial setup was handled through a temporary Wi-Fi network and the iOS application.
A photograph collection
The iOS application sent photographs and videos to the device. Haven stored the files and indexed photographs so they could be searched.
Search and interaction
A request from the app or voice interface was processed by Haven’s services. Mila searched the indexed photograph collection and returned matching results.
A shared display
The television followed the active agent. When Mila returned photographs, they could appear on the television as well as in the application.
The result was one system connecting a local computer, a phone and a television, rather than three separate experiences.
The wake-word listener and the television shared one small state file. The listener wrote which agent was active, what it was doing and what was asked; the television read it and followed.
The agent is one of mila, haven, roger or alfred. When a new query appeared, the television put it in its search box and ran it as a photo search. Switching agents by voice was added to this file on the day the first device left our household.
Limits
Haven 1 was a working first-generation system, not a finished consumer product.
- Setup and synchronization. The household-facing onboarding experience worked, but preparing each device required substantial manual engineering. New photographs and videos did not upload automatically as intended, and indexing new photographs remained an open issue.
- Search and voice. Mila’s search experience needed improvement, and wake-word recognition could be unreliable. Some place-naming functionality depended on an external geocoding service.
- Video and sharing. Videos could be synchronized and stored, but Mila did not display or play them. Sharing between separate households was not implemented, and the experimental television screensaver was not part of the delivered build.
- Release evidence. The original integrated implementation did not have a formal release tag, release-validation record or automated regression suite. Its deployment and household use demonstrated a working experience, not readiness for broad distribution.
History
Project 150
The original concept took shape around a device for the home, personal data ownership and AI agents. The early designs included family photographs, a television experience and the idea of software that could work across household devices.
Building Haven 1
We built the local device software, iOS application, photograph synchronization, Mila search, television interface, voice interaction and initial versions of the other agents. The original repository grew through 380 commits.
First external household
We deployed a Haven 1 device outside our own household. The recipient found onboarding and synchronization easy, while Mila’s search and wake-word interaction still needed work. The device was later returned.
A new implementation
We began developing HavenOS in a separate repository. The objective was to establish a more dependable, reproducible foundation for the software, while retaining the ambition behind the original household experience.
Revisiting Haven 1
We restored the original system to operation and revisited the integrated household experience it had already demonstrated.
HavenOS
Roadmap
Haven 1 and HavenOS are different implementations of the same broader direction.
Haven 1
The original integrated household implementation, including the device, iOS application, photograph synchronization, search, voice and television experience.
HavenOS
The newer software foundation, including the agent runtime, model interface and display contract. It does not yet reproduce the complete Haven 1 household experience.
Future development
Our direction is to bring the household experience together on a more dependable, reproducible and extensible foundation. This is direction, not an approved release commitment.