We built this for ourselves. That is where it started, and it is still the clearest explanation of why Mandaire exists and what it is trying to do.
Mandaire started as a private tool to solve a real problem. We built the first version for internal use. The problem was real: a large relationship network professionally and personally. A thousand conversations a year. Hundreds of threads with context that matters, deadlines that move, history that needs to be remembered. The kind of load that strains memory, attention, and time simultaneously.
AI tools helped with parts of it. Claude is remarkable at reasoning about a specific document in front of it. ChatGPT produces drafts with real judgment. Gemini synthesizes across Google's data. But each one knows only what it has been told in the current session, and each one lives in a separate silo. The result was a great deal of re-narrating: explaining to each AI who someone was, what the history between you was, what had been agreed, what was drifting. Every conversation starting from zero.
What was missing was not a better AI. What was missing was a persistent, encrypted, cross-provider layer that any AI could plug into and read. A context substrate that persists across all of them. That is what Mandaire is.
The first version ran for months for the people building it before any external user ever touched it. When it reached the point where a day without it would have been worse than a day with it, we knew it was real. That is the bar we apply to everything in it.
AI utility is gated by context. The better an AI understands your situation, your history, and the people involved, the better its output is. Every major AI company knows this. The difference is in how they respond to it.
Real-time access to your sources is now a commodity. MCP connectors, Apple App Intents, and native integrations mean any AI can reach Gmail, iMessage, or WhatsApp on demand. What that access produces is a shallow, stateless fetch: what happened today, in this app, in this session. That is useful. It is not the hard problem.
The hard problem is the resolved historical record. "Alex," "A. Rivera," and the person in your iMessage threads are the same individual. LLM memory cannot know that without being told, and being told once does not propagate across every source and every session. Deterministic entity resolution does, across every source, continuously. And resolution is only the first layer: the relationship graph, the calibrated inference, the disclosure rules that determine what to share with which person in which context, none of that exists in a retrieval system. It has to be built, maintained, and enforced upstream of any AI, before any query reaches it.
ChatGPT is the brain. Mandaire is what it cannot invent: a typed, time-aware, cross-source graph built from your full life corpus, with disclosure rules that run before any query reaches your AI. Who each person is, what your relationship with them warrants, and what you have decided the AI is allowed to say to whom. The defensible layer is not the connector. It is the trust-calibrated canonical store underneath it, and the recipient-aware disclosure engine that decides what leaves it. That is what we built, and it is a bet you can verify: the architecture is documented in full, the encryption model is public, and the data is exportable without friction at any time.
Gmail, Calendar, iMessage, WhatsApp, Apple Notes, AI conversation history. Read access only. We request the minimum scopes needed. You see exactly what we hold and can revoke any source at any time.
Your data lives on a dedicated server, not shared with any other user, never used to train any model. As we build toward full cryptographic privacy, we are implementing key-derivation architecture that places the decryption key on your device and out of our reach.
Claude, ChatGPT, Gemini. Mandaire connects to them as an MCP server. You do not switch AI providers to use Mandaire. You add Mandaire to the AI you already pay for.
Write tools (send email, add calendar events) require your explicit authorization per action type, then your confirmation on each individual action before it fires. Reading is automatic. Writing never is.
Your data exports in full on request. Not a truncated summary. The raw indexed material in formats you can read and re-import elsewhere. You own it. We are temporarily holding it for you.
Your information is used only to answer your questions. It does not leave your encrypted store to train models, inform advertising, or get shared with third parties. The business model is subscription revenue. Not data revenue.
Mandaire comes from mandataire: French legal term for one who acts on behalf of another, under explicit mandate, within defined limits, always in the interest of the principal. Never on their own account. The principal (you) retains authority. The mandataire acts on your behalf within the boundaries you set.
That is the standard every AI should be held to. Act on the user's mandate. Within the user's limits. In the user's interest. Not the company's interest, not the AI's interest, and not some algorithm's idea of what the user might want.
The logo is a weighted pair. A filled rectangle and an outlined one. The filled shape is you: present, load-bearing, the principal. The outlined shape is Mandaire: structured, capable, but defined by the boundaries you set. Different weights, one system. Neither is useful without the other.
The filled shape is you. The outlined shape is Mandaire. Different weights, one system. Neither is useful without the other.
Private beta. A small number of users are running Mandaire on both surfaces. We are onboarding carefully: personally, with direct setup support, and only after the product is ready for that person's specific situation.
If you want to use it: mandaire.app is for your personal life and professional relationships. mandaire.dev is for building products with your AI. Both have early-access request forms. We reply within a few days. No drip sequence.
If you want to talk about the underlying architecture, investment, or a partnership: [email protected]. We read it.
The full charter, architecture, and stated bets are at mandaire.org. If anything on any of our public surfaces contradicts what is written there, mandaire.org is right.