proactive agent · on a box that's yours
A personal chief-of-staff
that comes to you. Never sleeps.
Shimmer reads your inbox, calendar, and people. Thinks with your own knowledge base. Watches the rhythms that matter while you sleep — and proposes. You stay in command.
@you:shimmer.fyi
A different kind of assistant.
Same large models everyone else has — Shimmer routes to them. The difference is structural: proactive, contextual, naturalistic, and pre-tuned out of the box.
Comes to you, not the other way around.
Morning briefs, pre-meeting prep, follow-up nudges. The agent surfaces what matters before you ask. Measured by agent-initiated messages, not chat replies — if Shimmer waited to be opened, it would have no reason to exist over Claude.
Built from your context, not generic.
Persona extraction + a private knowledge base turn your existing comms into a structured model of you and your relationships. Every answer is grounded in it, with sources. The agent does the homework a chatbot makes you do.
Type what you want. No menus.
A Raycast-style command palette is the front door. Compound requests work: "draft a follow-up to Alex about Barbados," "remind me Thursday morning," "what is Mike saying about design system?" — the agent decomposes intent, drafts, awaits your nod.
Vertical. Zero setup.
Months of agent design shipped as defaults. Skills, rhythms, briefs, persona-extraction — already configured. You do not assemble a stack; the stack is the product.
The agent loop.
Learn · watch · propose · confirm. Repeats forever. One rule: if it ever asks you to approve something you didn't expect to be asked about — say no.
It learns you.
Reads your sent mail and meeting history. Builds a structured profile — who matters, what you care about, what you have decided. Stored in a git repo you own.
It watches for moments.
On a rhythm — morning brief, pre-meeting prep, periodic check-ins — the agent surfaces what is changing. Cold P0 contacts who need warming. Calendar prep. Follow-ups that slipped.
It proposes.
Drafts. Suggestions. "Schedule with Hira tomorrow if she is free." "Send this to Alex now or wait until evening?" The agent does the work and hands it to you.
You stay in command.
Everything destructive or external — outbound mail, calendar invites, third-party access — is gated. The agent asks; you confirm. It never sends without you.
Reads, classifies, summarizes, proposes. Sees your private data and untrusted inbound. Cannot send. Surfaces that handle untrusted content carry a faint amber rim — the boundary is felt, not announced.
Executes. Touches your accounts. Gated by TAP — the trusted-action protocol — so every outbound action requires explicit confirmation. The agent proposes; the user confirms; the action runs.
Pick a character. The discipline holds.
Your agent has a voice — you choose it. Pre-built personas to start from; customize freely. The tone changes; the gates don't.
The Biologist
Trusts observation over theory. Watches everything, says little, comfortable with the unexplained. Tells you what she sees — not what you want to hear.
Case
A console cowboy. Fast, instinctive, allergic to ceremony. Lives for the elegant cut-through and the direct line.
Roy
More human than human. Intense, acutely aware time is short. Stands just behind you and makes sure nothing that matters slips.
Murderbot
Reluctant, sardonic, fiercely competent and protective. Would rather not talk to you, but will keep you out of trouble and call a bad idea a bad idea.
Your data. Your control. Honestly framed.
Three things are true right now. One thing is on the roadmap — and we will not pretend otherwise on a marketing page.
Your knowledge is yours.
A git repo you own. Personas, decisions, context, history. Portable; take it with you if you leave.
Outbound is gated.
TAP — the trusted-action protocol — fires on anything destructive or external. The agent proposes; you confirm; the action runs. Never autonomous on outbound.
No model trains on you.
Your data is not used to train anyone’s model — yours or anyone else’s. Inference passes through; nothing sticks.
Full self-hosting on your hardware.
Today, Shimmer runs on a shared VPS (operator has host access; LLM inference passes through Anthropic / OpenAI / similar). The single-tenant container architecture means full self-hosting is achievable — and it's the obvious next step. We'll tell you when it ships, not before.
Request your personal agent.
Ten minutes from request to a working agent that knows your world, watches the rhythms that matter, and works on your behalf — without ever sending anything you didn't see first.