HarperHead of HR · your AI recruiter

Harper is AI — she's here to help you recruit your next AI employee and build your agentic company. Files you attach are read in your browser, never uploaded.

HarperAsk anything — your guide to the portal, the catalog, and your team

Harper is AI — she drafts, you decide.

Sessions

Every session with Harper — each one a hire in progress. Reopen to keep building, rename, archive, or delete. Your team and briefs save to your account and follow you across devices; the conversation text stays in your browser.

Company context — carried into every new session

Show archived

Quick start

The full castEveryone you can hire — the roster you build from

Open full ↗

Your company, at a glance

You're the CEO. ATLAS, your Chief of Staff, runs the day-to-day. 10 named leaders each run a department, and behind them sits a bench of 140+ specialists you can hire onto any employee.

1 Agentic CEO · 10 department leaders · 140+ specialists

Catalog

Every skill, grouped under the role that owns the outcome. Each one is its own page with a shareable link — click through to read, copy, or send it. Open the full library ↗

Ask HarperHead of HR — describe the outcome and she matches the skills + the role that owns them

The Brain protocol — click the nucleus to download

Your Brain — the Company Brain Atlas Folder

The knowledge unique to you — frameworks, clients, how you work — in one place. Every AI employee reads it first, so the work sounds like your company from draft one, and grows as you do.

What's inside — the folder structure

One download, seeded from your session brief. Front door at the root, everything else organized under Company/.

soul.mdidentityRead first. Who you are, your voice, your hard boundaries — owner-gated, changed only with you.

company-brain.mdfront doorYour Chief of Staff's charter and the map of everything else in the folder.

CLAUDE.mdpointerThe thin boot pointer any tool auto-reads first — where your brain lives, the boot order, the hard gates.

Company/The shared tier: skills, SOPs, playbooks, roles, the FAQ inbox, connectors, access rules — and Clients/ if you serve more than one.

AI-Team/Your AI employees — a desk each, added as they're actually hired. No pre-built roster.

People/One desk per human teammate — their calls, their daily notes, their charter.

Daily-Notes/The ledger — one dated file per working day, written live; every session ends with a Session Close.

Archive/Superseded material — never deleted.

Private/Owner-only. Payroll, banking, contract terms — never inside a team-shared Drive.

One zip, one prompt: unzip it, connect your MCP, paste the prompt into Claude inside the folder — it asks whether you're starting blank or harvesting from real docs, then runs the setup.

How it runs — the deployable model

Not a filing cabinet — a working system, field-proven on Drive + Claude. Across seven layers your AI staff captures, processes, and distributes the work, keeps it cheap, and governs and improves itself. You bring judgment and the go-ahead.

CaptureMeetings, calls, and videos land automatically — recording, transcript, chat, participants, filed by date. The raw capture is immutable truth; everything downstream derives from it.

ProcessA Knowledge Agent reads the raw files on a schedule and files recaps, action items with owners, decisions, open questions, and profile updates. Its behavior is a plain-markdown skill file you edit live — no deploy.

ReviewA graduated gate: deterministic checks first, an independent AI reviewer second (rejecting privacy leaks, wrong owners, guesses stated as fact), a human only on the consequential — privacy, commitments, money, anything customer-facing.

DistributeYour function agents act as real employees: a Project Manager runs the tracker and standups, your Chief of Staff and Librarian keep the record and the digests, a Director routes the fleet and owns deploys.

CostFour-layer caching, a model router (cheapest capable model per job), and one cost monitor under a hard spend cap keep the whole brain running for a few hundred dollars a month — tracked to the dollar, alarmed before the wall.

AutonomyThe brain improves itself, safely: an improvement agent files ranked tickets, a builder works by pull request only, an independent reviewer gates the merge, and a failed deploy auto-reverts. Protected paths stay human-merge-only, always.

GovernanceTwo keys, two powers (a read key for snapshots, an act key held tight), heartbeats so silence is never ambiguous, and gates that live in your SOPs, not anyone's memory. Any human or AI builds the same way.

Rolling it out — crawl, walk, run

Don't build it all at once. Each phase earns trust before the next — value on day one.

Week 1 · Crawl

Fill or harvest

Your soul confirmed, your Chief of Staff named and chartered, one Knowledge Agent producing recaps. Value on day one.

Weeks 2–3 · Walk

Operate

The daily journal close and call-filing sweep scheduled, your first specialist hire on the bottleneck — and the cost monitor from day one, before the bill surprises you.

Week 4+ · Run

Re-architect + prove

The re-architecture pass, the negative test, and permissions for the team — turned on only after the basics have earned trust.

The whole stack: Google Drive, Slack, a Claude API key with a spend cap, and GitHub only when you turn on autonomy. No new project-management tool, no middleware.

The protocols — the habits that keep it alive

Boot order, every session. soul.md → company-brain.md → Company/ → the last few Daily-Notes/.
The ledger. Daily-Notes/ is written live as work happens, never batched — every session closes with a Session Close: what was logged, what changed, what's pending, the next one thing.
A rule earns the shelf at the second occurrence — never the first, never "someday." One-offs just stay ledger lines.
Captured is not permission. Nothing sends, publishes, or moves money without your explicit go.
Delete superseded protocol — never stack. Edit in place and log the change; stacked "this wins" blocks are how a brain stops parsing itself.

Build SOPs — the operating discipline

The reusable rules for running an AI team safely and cheaply — the same discipline we run ourselves, generalized for you. They ship inside your Company Brain Atlas Folder at Company/SOPs/, and you can grab any one here, or the whole set.

Start here — the index + 7 principlesThe map, the kickoff prompt, and the seven principles that govern everything.

API Cost Discipline (caching + routing)The five rules that keep the API bill sane.

Caching Check Gate (six checks)The six checks every LLM job passes before it ships.

Building a New Agent (+ Autonomy Gate)Identity, wiring, and the gate for agents that can act on their own.

Skill OnboardingAdd a capability without breaking the fleet.

The Docs + Skills GateSync the skill files, not just the docs — the #1 recurring failure.

Security: Two Keys, Two PowersRead power vs act power, and where secrets live.

Ship DisciplineRecord everything shipped; close every loop.

Automating Your Skills (Drive + Claude)The skinny-launcher registry: one canonical skill, planted everywhere, drifting nowhere.

Google Drive Conflict Copies (drift)A suffixed file is never your work — the STOP/fold/archive rule, and the git-in-Drive guardrail.

Who sees what

Enforced by Google Drive sharing, not a separate permission system — each person's AI inherits exactly their own access.

Who	Sees	Why
Owner	Everything, including `Private/`	The one seat with no walls.
A client/location manager	Their `Company/Clients/<theirs>` + `Company/` (read)	Everything a deal or relationship touches — not the other locations.
Floor / new hire	`Company/Playbooks/` (read) only	What they need to do the job, nothing they don't.
Chief of Staff	Everything	The integrator — routes work across the whole map.

Local or Drive — same folder either way. Plain .md, zero dependencies, works local-only, nothing to install. Prefer Drive? Put it in My Drive (or a Shared Drive for a team) and point Google Drive for Desktop at it set to Mirrored ("Available offline") — a streamed-only file shows your AI employees just a stub, not the real content. Obsidian is a viewer, never a requirement — nice if you like it, strictly optional.

The capture stack — feeding the brain

Meeting recorders are the fastest way in. Google Meet's built-in notes is the zero-code path — no new tool, no account. Zoom, Fathom, and Granola work too. The call-filing-sweep skill (one of the two that ship scheduled in Company/skills/) files each recording by nature — 1:1s to a person's desk, cross-team calls to the shared Transcript Vault — no manual inbox step.

Deeper Google access (advanced)

Want your AI employees reading Gmail, Calendar, and Drive across more than one Google account? Install Gmail Multi-Account Connector — a separate Google Workspace MCP server per account, each with its own credentials. Every key stays with Claude Code; none ever live in a file.

It's yours — not ours. The Company Brain Atlas Folder lives on your machine and Drive, in plain .md files you own. We hold nothing — no copy, no transcripts, no lock-in. Walk away any time, whole brain intact.

How your stuff is handled

Your business brain stays on your machine. Harper borrows it to think — she never keeps it. Seven promises, no fine print. Account data is named separately below so "not stored" never overclaims.

Read, never kept. Your files are read in your browser and used to ground the conversation you're having. The relevant text rides along with each question to the AI services that answer you, and it's gone from our side — we don't store it, we don't log it, and per our AI providers' commercial API terms it's not used to train models. (Providers may retain data briefly for abuse prevention under their own policies.)
Your brain lives on your machine. The Atlas Folder is plain markdown you own. Sync it with your own Google Drive if you want. Cancel tomorrow and you lose our catalog — not your brain.
Your account data is limited and explicit. Supabase stores login data, your private profile, profile-sharing choices, and the structured hires and saved skills that follow you across devices — including a short, capped list of specific facts Harper carries forward so she doesn't make you repeat yourself, never the conversation itself. A GHL onboarding survey can prefill that private profile for review; it cannot publish you. Directory receives only the display fields you approve. We do not put your email, access level, goals, bottlenecks, CRM notes, or Company Brain content in Directory.
Your conversation text lives in your browser. We sync the structured brief, hires, and saved skills needed for your account, not the Harper transcript or attached-file text. Export any session to your Atlas Folder as a plain .md — one click in — so nothing irreplaceable lives only in a browser.
Your voice isn't recorded by us. The mic drives the orb animation and your browser's own speech-to-text (which may use your browser vendor's speech service) — we receive text, never audio. Harper's voice is generated by ElevenLabs from her reply text.
We only look where the public can. The site scan reads what any visitor to your website could read. Nothing behind a login, ever.
We tell you who's in the flow. Anthropic handles Harper's thinking · ElevenLabs handles optional voice · Firecrawl reads a public site scan · Vercel hosts the app · Supabase handles login and account context · GHL holds membership and onboarding intake. If this list changes, this page changes first.

If you leave

Keeps working forever	Stops with membership
Your Atlas Folder and everything in it · downloaded install packets and starters · exported sessions. Plain files, licensed to you — we never revoke, expire, or phone home.	The gated skill catalog (the runtime your AI employees draw on) · Harper and this portal.

What we don't promise — on purpose. No certification cosplay ("enterprise-grade"). No accuracy guarantees — Harper is AI: she drafts, you decide, verify before acting. And no frozen vendor list — we promise to update this page first, not to never change.

Share responsibly. Bring context you have the right to share. Don't paste other people's sensitive personal data — your clients' trust is your Atlas Folder's foundation too.

Updated July 2026 — account-profile and Directory privacy pass. This page describes what the product actually does; when the product changes, this page changes with it.