The first step takes two minutes. After that, every conversation with Claude, ChatGPT, or Gemini has access to everything you choose to give it: who people are, what was agreed, what is about to slip. Depth compounds from there.
There is no prerequisite. You do not need to import anything before the product is useful. The entry point is installing Mandaire on your AI and using it as you normally would. Everything after that adds depth.
In Claude, ChatGPT, or Gemini settings, add a new connector. The URL is the Mandaire MCP server. One authorisation handshake. After that, your AI and Mandaire are connected and stay connected. Use your AI as normal.
Export your ChatGPT, Claude, and Gemini conversation archives and upload them. Mandaire indexes the reasoning you have already done with AI tools. The prior conversations become part of the context across every provider you use.
Gmail and Calendar give Mandaire your actual communication and scheduling history. Each connection requires an OAuth authorisation in your browser. Minimum access only. Read-only.
iMessage, WhatsApp, Apple Notes. These add the conversations that never made it into email: the real exchanges, the context your inbox does not capture. Each source is the same pattern: OAuth authorisation, read-only scope, local indexing.
Mandaire is an MCP server exposed over Streamable HTTP with OAuth 2.1 and PKCE. The AI provider performs a standard OAuth authorization code flow at mcp.mandaire.com. After authorization, the AI sends tool calls directly to the MCP server in each session.
AI conversation history imports are handled via a web upload endpoint. You export a zip from the provider, upload it, and Mandaire indexes the conversations locally before encrypting the result. The raw export file is not retained after indexing.
Sources are ingested via their respective OAuth APIs (Gmail API, Microsoft Graph, WhatsApp Business API for personal use, EventKit for Calendar/Reminders, iCloud API for Notes). OAuth tokens are stored encrypted alongside the synthesis index. Mandaire holds tokens with read-only scope. Token refresh is automatic.
The difference is context arriving without re-narrating it. You do not explain who someone is. You do not paste an email thread. You do not say "we discussed this last month." Mandaire carries that. The AI reads it when it needs it.
The AI does not just know more. It knows the right things at the right moment. A question about a meeting brief pulls the person's history, the prior conversation thread, and anything outstanding between you. A question about a draft email knows the recipient's communication style from two years of actual messages, not a summary you wrote.
When you send a message to an AI with Mandaire connected, the AI calls the mandaire tool with a verb and a data kind, for example mandaire(verb=SELECT, from_kind=person) to retrieve a contact's relationship history, or mandaire(verb=SELECT, from_kind=email) to search your inbox. These are read-only calls to the Mandaire MCP server. The server returns structured context that the AI incorporates into its response.
The AI decides when to call Mandaire based on the shape of your question. You do not instruct it to. Questions that name people, reference past events, or require relationship context trigger Mandaire lookups automatically.
The synthesis layer Mandaire queries is an encrypted entity and relationship graph, not a raw message store. The AI receives resolved entities (people, commitments, threads) rather than raw email bodies. This is faster, more private, and more accurate than passing raw text.
Two examples of the kind of question that becomes answerable.
You had a conversation three weeks ago with someone you are meeting tomorrow. You remember the general topic but not the specifics, and you did not keep notes.
The AI did not know any of this before the query. It read it from Mandaire as part of answering your question.
You are about to build a feature and you have a nagging feeling you already decided against this exact approach once before.
The prior reasoning was in a ChatGPT conversation, which Claude has no access to. Mandaire does, because it indexed your AI history across all providers.
Neither example requires a specially formatted query. Both work because the AI can reach Mandaire when the question naturally calls for it.
The first session, Mandaire has only what you gave it before you started. The AI is better than it was, but the gap is small. The context it returns is shallow.
By the fiftieth conversation, Mandaire has seen how you think, who you mention, what you keep asking about across different topics. The entity graph has resolved the people you interact with most: who each person actually is across every source, what they care about based on years of interaction patterns, what your relationship with each of them looks like, what you typically share with each, and which claims about them across your corpus are reliable enough to act on. Commitments have been tracked through resolution. Decisions you have made with other AIs are now available to this one. The AI is working with context that does not fit in a context window and never existed in one place before.
By year one, the synthesis layer knows patterns you would not know to ask about. A person's communication cadence changing. A topic that keeps coming up in different forms. A commitment that was restated three times without resolution. These are not things you would think to put in a note. They emerge from the index.
This is why the product is positioned as a substrate, not a feature. Memory products that store summaries are good for recall. A synthesis layer that compounds over time is useful for judgment. The two are structurally different products.
This is an architectural constraint, not a privacy policy. Your synthesis index lives on your hardware, or on the dedicated server you designate in the managed tier, and does not leave it. No one on the Mandaire team can read it through any normal operational channel: no admin panel, no support tool, no bulk access path.
We are building toward a stronger guarantee: key-derivation architecture that will make it architecturally impossible for anyone, including us, to read your data even under legal compulsion. That architecture is under active development and not yet complete. When it ships, a government request or infrastructure breach will produce encrypted data that no one can read. We will update this page when it ships.
This is also what makes Mandaire inspectable and correctable in a way that hosted AI products are not. The full entity graph, the relationship index, the commitment ledger, the AI conversation history: all of it is in files on your machine, in formats an engineer can read without Mandaire's involvement. If something is wrong about what Mandaire believes about a relationship or a decision, you can correct it. If you want to leave, you take the index with you. The export format is documented and does not require us to cooperate.
Mandaire uses widely-audited open-source encryption libraries. The architecture is described in full at mandaire.org.
The most common question before connecting the first source: how long before this is useful? The answer depends on what you mean by useful.
Ask your AI who you have been emailing most in the last 30 days. Ask about the last thread with someone specific. Ask what is on your calendar tomorrow. The index is shallow but real. The gap from a general AI is already visible on the first query.
Threads that went quiet. Something you committed to that nobody followed up on. A meeting on your calendar you have not prepared for. This is when the useful-by-accident moments start.
Not because it is clever about writing style. Because it has seen two years of how you actually write to each person, and your calling AI can use that context directly. The draft sounds like you because it knows more about the relationship than you had time to narrate.
Cross-source patterns. A topic you keep returning to in different forms. A relationship whose communication cadence has shifted. These are not things you would write in a note or put in a search box. They emerge from an index that has been running long enough to notice them.
If week two still feels like a generic search engine, reply to the onboarding email. It usually means a source did not connect fully. Do not quit before we have looked at it together.
Mandaire is in private beta. We onboard personally: you connect your sources with direct support, run a first brief together, and check back at day seven. We do not send you a link and wish you luck.
The full architecture is at mandaire.com/architecture. The charter, stated bets, and cryptographic specification are at mandaire.org.