Haven is a self-organizing ontologyA structured model of the concepts in a domain and the relationships between them. Borrowed from philosophy; used widely in AI and information science. designed for human ↔ agent collaboration.
// Private. Open Source. Deploys anywhere.
Alignment tracks the data you align against;[1][2] data quality dominates model choice.[3] The ceiling on how useful an AI can be for you is the same as the ceiling on how much of your life it can safely see. To break through it, an agent has to reach the things you'd normally keep private.
So how do we safely give agents access to our private data?
Handing it over to OpenAI, Google, or Anthropic isn't a real option — hopefully most people already know better. Doing so would be handing over every intimate detail of your life willingly, to a company whose incentives, ownership, and policies have proven time and time again to be adversarial to its users.[4][5][6][7]
The other option is Haven. Haven runs in an environment you either trust or control. It provides the storage, security, and query primitives an agent needs understand your ontology, act on your behalf, and get better over time. No matter what, you stay in control.
A private cloud is infrastructure dedicated to a single person or organization — as opposed to a public cloud like Google Workspace or iCloud, where you share compute, storage, and often data-processing pipelines with millions of other users.
Haven runs your workspace on hardware you own, or on a single-tenant instance we run for you. The data, the AI queries against it, and the derived indexes belong to you, not to the platform.
Google Drive is a file host. iCloud is a device-backup service. Haven is a workspace that ingests everything you produce or receive — mail, calendar, files, photos, notes, transactions — into a single graph that humans and agents both query.
And unlike those companies, Haven's business is running the software, not monetizing what you put in it.
The ontology is the graph of everything Haven knows about your data. Every email,
event, file, note, and transaction is a node, connected by typed edges — author,
participant, mentioned, attached-to, and so on.
Self-organizing means Haven creates and maintains those links as data arrives. You don't design a folder tree or hand-tag anything.
Any provider that speaks the OpenAI or Anthropic API — that covers OpenAI itself,
Anthropic, Google, and every open-weights runner (ollama, vllm,
lm-studio, etc.).
You configure the endpoint and pass credentials through Haven. For sensitive data, point Haven at a local model that never leaves the machine.
To some extent, yes. What's different is the business model. We charge you for running the software, not for access to your data. We don't sell it, train models on it, or hand it to advertising networks.
And you can move to a self-hosted install once that ships — so trust in us is bounded, not permanent.
After the hosted product stabilizes. Internals change most weeks in alpha and running a self-hosted install against a moving target would be painful for everyone.
No fixed date. We'll announce a target once the API is stable enough that a self-hosted deploy would survive a month without breaking.
On hosted: only what you put in, plus the operational metadata needed to run it (server logs, billing records, error traces). No third-party analytics, no fingerprinting, no ad SDKs.
On self-host: none of the above leaves your machine, ever.