Operationally
Before you connect twenty years of email and a decade of messages, understand what the pipeline actually does with it.
The pipeline
Your data is pulled from your connected sources, Gmail, iMessage, WhatsApp, Calendar, ChatGPT and Claude conversation history, and indexed in an encrypted store that lives on your hardware (or ours, if you use the managed tier). A local reasoning model runs against that store continuously. Not the frontier model you pay for elsewhere; a local model that handles the routine work: building your relationship graph, tracking what you have committed to, flagging what is going quiet.
Your AI client connects to this via a single MCP tool called mandaire. When you ask Claude or ChatGPT something in your normal chat interface, it retrieves the relevant context from your store and passes it to your client. The frontier model never has access to your full archive. Only the context relevant to the specific question is ever surfaced. A question about one person does not surface a different person's information.
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 zero-knowledge primitives that will make it architecturally impossible for anyone, including us, to read your data. The privacy page covers the current state and what ships next.
What compounds
The first week, Mandaire can search your email and calendar well. Your morning brief is accurate. If you ask about someone, it returns the thread. This is useful but not yet remarkable; it is roughly what a well-configured search tool does.
By month 3, the picture is dense enough to surface things you have not thought to ask. Who has gone quiet that usually does not. What you said you would do and did not follow up on. Which relationship is drifting in a direction where a note now, rather than in two weeks, lands differently. The gap between what week 1 feels like and what month 3 feels like is real. It does not come from the AI getting smarter; it comes from the data getting richer.
This is the reason the onboarding sequence starts with AI conversation history first, then email, then messages. The earlier exports (ChatGPT, Claude, Gemini) are often three to five years of distilled thinking, the fastest way to seed a picture. Two gigabytes of structured export installs in minutes. Email takes hours but covers twenty years. Messages take days but cover the relationships that never moved to email.
The LLM stack
Most personal AI tools route everything through a single frontier model, Claude API, OpenAI API, and pass the cost directly to you or roll it into the subscription. Mandaire does not do this, because the per-call cost of running frontier inference against a full personal archive at the rate a useful daily tool requires would be prohibitive.
A local model handles the routine work continuously: brief generation, relationship-graph maintenance, commitment tracking, decay detection. This runs against your store throughout the day at no per-query cost to you after the flat monthly fee.
Claude Desktop, ChatGPT, Gemini, whichever you already subscribe to. Mandaire exposes your context to it via MCP. You do not need a new chat surface or a new subscription. The setup is one step: connect your existing client to the mandaire tool.
We never see your Claude Max or ChatGPT credentials. We never resell tokens. Your renderer subscription is yours; Mandaire is the context layer the renderer reads.
Pricing and sovereignty
$50/month (beta)
We operate your dedicated store on hardware we control. Your data is not shared with any other user. For people who want the system working without operating infrastructure.
Free (open source)
Self-host the full stack on your own machine or cloud account. Right for people with a home server, a spare Mac mini, or a preference for running their own infrastructure. Managed and Yours are not a quality split.
You can export your full archive and migrate to self-host at any time. No lock-in is an architectural commitment, not a preference. The encryption module is open-source under AGPL, which means any competent engineer can read your archive without Mandaire's involvement.