A personal Office of the CEO

An AI partner that knows you, and is always thinking about you.

Not an assistant that waits for instructions. My Harness builds a living model of your world that evolves and learns, so it keeps you current on the people, commitments, and priorities that matter.

private per person · built for teams works with any AI your data, your files status:
The objective
Deliver AI as a true partner: one that knows you deeply, is constantly thinking on your behalf, and earns your trust by being right and by knowing when to stay quiet.

The problem it solves

The gap isn't that AI forgets. It's that nothing actually knows you, or works for you, between the moments you ask.

01

AI doesn't know you

Every session starts from zero. The assistant is a stranger each time, never building a real model of you, your people, or what you're trying to do.

02

Staying current is invisible work

The real tax isn't forgetting, it's the constant labor of reloading context before every meeting, reply, and decision.

03

Tools wait to be asked

They're passive and reactive. Nothing is thinking ahead for you, anticipating what's coming, or catching what you'd otherwise let slip.

What we're building: a partner, not a tool

Three promises, and the discipline that makes them real.

It knows you

A rich context layer that compounds, your relationships, commitments, and goals, sharpening every week and following you across whatever AI you use.

It has your back

It keeps you current and catches what you'd drop, so you walk in prepared and never fall behind, without doing the upkeep yourself.

It looks out for you

Always-on and thinking on your behalf, anticipating what's coming and surfacing what serves your goals, in your interest.

Earned by restraint. A partner that's constantly thinking about you has to know when to stay quiet. It thinks all the time and speaks rarely, only when it's worth your attention, and it measures itself on exactly that.
Why it matters

Time back, nothing dropped, an asset that compounds.

It is measured, not asserted. The system grades itself every week on net attention, the time you give it versus the attention it gives back, and publishes the number.

0
net attention returned in real use, to date (handled-for-you minus asks)
net-positive
the everyday loop now returns more than it asks, week over week
hundreds → single digits
open items waiting on you, once the self-cleaning loop took hold
01

Time returned

Routine prep, follow-through, and bookkeeping handled in the background, proven by the net-attention meter.

02

Nothing drops

Every commitment, follow-up, and cooling relationship is tracked and surfaced before it slips.

03

Show up sharper

Walk into every meeting already briefed, and know which relationships are warming or cooling, with the reason why.

04

An asset that compounds

The moat is a longitudinal record of how you operate. After years of continuity, switching cost becomes the relationship itself.

05

From personal to company

The same private foundation extends to a team: a shared, trustworthy company brain where each person's data stays their own.

Capabilities

Memory is table stakes. State is the product.

Most tools store and retrieve. My Harness adds the two layers above: a live model of where things stand, and a read on what changed and what to do about it.

01 — WHAT HAPPENED

Events

Email, calendar, meetings, health. The record everything traces back to.

02 — WHERE THINGS STAND

State

Relationships, commitments, readiness, each value with its own confidence.

03 — WHAT CHANGED

Understanding

The deltas feed the brief and the alerts. Noticing change is the point.

01

A morning brief

One page each day: meetings, who you're seeing and what matters about them, live follow-ups, readiness. Assembled overnight, on its own.

02

Relationship radar

Trust, momentum, and risk per person, recomputed daily, flagging who is cooling before you'd notice.

03

A thinking partner

How would this person respond? Who should I talk to about X? Grounded in real history, with citations.

04

Meeting intelligence

Every transcript becomes structured signal: decisions, roles, candor, and a rolling read on each person.

05

Memory that compounds

Daily rolls to weekly, monthly, and yearly, plus an autonomous pass that surfaces connections you didn't ask for.

06

Keeps itself current

It closes follow-ups you've handled, ages out the dead, and drains its own questions, so the list stays short.

07

Chief-of-staff drafting

It drafts the memo, brief, or reply from context. Nothing leaves without your approval.

08

Teams and shared projects

A private instance per person, and a shared company layer, each person's data still their own.

Architecture

How it fits together.

Your machines hold data and feed it. The thinking happens in the cloud LLM, so the system stays cheap and works with whatever AI you use. The model never runs on your hardware.

My Harness — how it fits together You + any AI client Claude · ChatGPT · Cowork + web dashboard LLM — the thinking Anthropic — Sonnet + Opus model-agnostic · not on your hardware Your Mac only when it's open LinkedIn scrape (Chrome) iMessage + dropped files Cowork enrichment agents VPS — the core always on FastAPI + MCP server your data: JSON files, encrypted memory · state · semantic index nightly jobs: cascade + brief · no model runs here Cloud sources via OAuth / API Gmail · Calendar Plaud transcripts Oura · Withings OAuth / MCP app.myharness.ai its own LLM daily brief push scrape API pulls Your machines hold data and feed it. The thinking happens in the cloud LLM, so the box stays cheap and model-agnostic.
Topology — what lives on the VPS vs your Mac vs third-party cloud.
Data model

Events to state to understanding.

Raw evidence becomes a live model of your world, which becomes an understanding of what changed and what to do. All of it plain files, no database.

My Harness — the data model events → state → understanding · all plain files, no database EVENTS — the evidence (raw, append-only) Email · Calendar · Meeting transcripts · LinkedIn · Health (Oura/Withings) · iMessage STATE — the live model People / CRM profiles · facts · warmth Organizations boards · communities Projects workspaces Meetings transcripts → intel linked by slug — the people-graph · each = profile + facts + state The State Object — the shared primitive dimension = value + confidence + evidence deltas = what changed, with the “why” history = prior snapshots (the time series) provenance · inspectable · editable e.g. trust 4.2 → 3.1 (cooling) UNDERSTANDING — what changed + what to do Memory cascade daily → weekly → yearly Deltas what changed this week Dreams + insights autonomous synthesis Delivery brief · standing context · search Filesystem-JSON · one folder per user · git-backed · encrypted at rest · every change reversible · no database
The three layers, built on the State Object primitive.
Under the hood

Boring infrastructure, on purpose.

A small set of Python modules and flat files on one inexpensive server, exposed to any AI client through open standards. Context lives outside the model, so it is not a model wrapper.

transportFastMCP + FastAPI. A scope-gated MCP surface plus a remote streamable-HTTP transport. Any MCP client consumes the same externalized state.
authOAuth 2.1 with dynamic client registration plus per-client scoped API keys. Every tool maps to a required scope, enforced before dispatch.
storageFilesystem-JSON. No database. Every belief is a plain file, git-backed, diffable, inspectable, and reversible.
state layerState Objects. Dimensions with confidence and evidence, first-class deltas, and append-only history. Change detection is the product.
retrievalSelf-hosted semantic index. A local CPU embedding model over the memory tree; no corpus text or query ever leaves the box.
cost modelTools never run an LLM. The connected client composes with its own subscription, so operating cost is near zero and predictable.

Entity relationship model

The logical model. It's filesystem-JSON, so these are conceptual entities, references are slugs in files, not join tables. PERSON is the hub; the State Objects (RelationshipState, CommitmentState) carry the confidence-and-deltas shape.

My Harness — entity relationship diagram logical model · stored as plain JSON files (no join tables — references are slugs) 1N 1Nowns N1alias of 11scores N1has N1owed NMmembers MNinvolves MNattendees MNtagged MNinvolves N1from N1spoken by TENANT PK tenant_id FK members[] → User USER PK user_id role owns every row (isolation) IDENTITY_ALIAS PK alias_slug FK canonical → Person collapses duplicates RELATIONSHIP_STATE «State Object» PK id (= person slug) type = relationship dimensions: trust / momentum / risk value + confidence + evidence deltas[] (what changed + why) history[] (prior snapshots) recomputed daily · deterministic REMEMBER_FACT PK id (content hash) FK person_slug → Person text · category importance · confidence relevant_until · status family / work / health / milestone PERSON PK slug profile_url (unique) full_name · display_name current_company warmth · context project_tags[] last_interaction FK owner → User OPEN_LOOP «Commitment State» PK loop_id FK person_slug → Person description owner (me/them) · status source · follow_up_date open → closed / superseded ORGANIZATION PK slug type (board/cohort/community) status · summary FK members[] → Person observations[] append-only PROJECT PK slug status · keywords[] summary · progress[] FK people[] → Person FK orgs[] → Organization links[] MEETING PK filename date · meeting_type effectiveness FK participants[] → Person FK projects[] → Project decisions · seeds SEED (latent idea) PK idea_slug FK speaker → Person FK source_meeting → Meeting related_people[] PK primary key (slug/id) FK reference [] list of slug refs «State Object» = value+confidence+evidence + deltas + history Cardinality: 1 = one, N/M = many. References are slugs inside JSON files — there are no join tables; the “graph” is computed over the files.
Logical entity relationships — PERSON is the hub.

Safe to rely on, by construction

Per-user isolation

Multiple enforced layers; cross-user leakage fails the regression suite.

Reversible by default

Every automated action writes its inverse to an append-only journal. Undo anything.

Delete-proof autonomy

Nightly agents hold scoped keys that cannot reach destructive operations.

Prompt-injection boundary

Untrusted text is wrapped and quarantined before any model reads it.

Encryption at rest

Credentials encrypted under a master key. No training on your data.

Yours to export

The corpus is plain files you own. No lock-in.

No database Self-hosted embeddings Model-independent (MCP) Git-backed + inspectable Private per person · built for teams