It started with you, not a rankings list.
Carabase surfaced durable signals from past meals: seafood and produce, serious wine, a love of Mugaritz—and enough disappointing Arzak visits.
The personal context engine
Carabase turns the useful signal across your mail, messages, photos, notes, calendar, code, health and media into private, living context for the AI tools you choose. It runs on hardware you control.
YOU
What does my actual listening say I like right now—not what I think I like?
CARABASE
You still describe yourself as a nu-metal and electronic person. Lately, though, you have been living in Noah Kahan, Justin Bieber, Sam Fender and Olivia Dean. Your taste has not narrowed. Its centre of gravity has moved.
A real Carabase session
Codex used two layers: what you actually like, and what was genuinely open and relevant next week. We sped up the waiting; the work is real.
Carabase surfaced durable signals from past meals: seafood and produce, serious wine, a love of Mugaritz—and enough disappointing Arzak visits.
Codex checked current menus, prices, opening days and reservation reality. Recent diner reports were practical clues, not truth.
Fit, distinctiveness, scarcity, availability and a credible fallback. Elkano and Rekondo rose; another Arzak visit fell away.
The result was not another “best restaurants in San Sebastián” list. It was a shortlist that could only have been made for you.
The difference
Project context lives across code, issues, agent sessions and conversations. Carabase keeps the decisions attached.
The repository reveals current behaviour. It does not contain every rejected approach, unresolved trade-off or safety boundary.
The context problem
People should not have to paste their life into a chat window—or maintain a perfect second brain—before AI can help.
Noah Kahan leads your recent listening by a wide margin.
Retrieval changed. Recall widening remains default-off.
Tastes, projects and plans—connected, current and cited.
The commitments that make advice useful right now.
What you said, decided and learned along the way.
How it works
Three moves turn an impossible pile into useful, inspectable context in the moment it is needed.
Authorised connectors and on-device harvesters collect signal without asking you to file, tag or remember the right folder.
MAIL · MESSAGES · PHOTOS · NOTES · CALENDAR · CODE · HEALTH · MEDIAEntities, relationships, time and provenance become Folios: cited context about the people, projects and tastes that matter.
EVERY FACT KEEPS ITS ORIGINSemantic, lexical and graph retrieval return grounded context with resolvable receipts—or abstain when the evidence is thin.
ONE CONTEXT LAYER · EVERY ASSISTANTLife-scale context
Recommendations shaped by what you play, what you own and how your taste changes over time.
What changed, why it changed and which assumptions the next agent must preserve.
Patterns across activity, recovery, appointments and personal notes without a cloud life archive.
Model choice
Carabase is the neutral layer between the life you already live and every assistant you choose.
Mail, messages, files, photos, calendars, code, health, media and more.
Time, entities, relationships, provenance, Folios and retrieval on hardware you control.
Everything relevant. Nothing surrendered.
Privacy is a property of the architecture, not a promise buried in a policy. Carabase keeps storage, transport and inference boundaries explicit.
Working now
Carabase already runs our lives, and we can’t live without it. Now we’re looking for a few more devotees.
Life-scale signal, not a folder of curated files.
The full loop can run against a model the user controls.
Daily Log, Folios and Summon; iOS companion in development.
Corrections become future retrieval signals.
The open-core promise
The host—where all your data lives—is open source, and always will be. Don’t take our security claims on faith: inspect the code for yourself. If you’d rather not tinker, a hosted version is coming, built on exactly the same core.
Hosted version coming · same open coreThe useful objections
The short answers. The deeper thesis and technical detail remain available for people who want to inspect the argument as closely as the code.
A persistent layer between your digital life and the AI tools you choose. It gathers authorised signal, turns it into living, cited context and retrieves the relevant pieces when an assistant needs them.
Your source data and Carabase knowledge substrate live on your hardware. The host exposes no public ports. If you choose a cloud model, Carabase makes the context leaving your machine for that request explicit.
Yes. Carabase is the context layer, not another chat window. Codex, Claude, OpenClaw and local models can all use the same grounded context rather than building separate memories of you.
Yes. The host—the part where your data lives—is open source and will remain so. A managed version is coming on exactly the same core for people who value convenience over tinkering.
The private alpha is for technically confident people with a recent Apple Silicon Mac. We will do the setup together, then ask you to use Carabase on real work and tell us exactly what delights, confuses or breaks.
Private alpha
I am looking for technically confident people who use AI heavily, care about privacy and will try Carabase in real life—not just admire the architecture.