Highside AI Concierge (Cooper)

Highside AI Concierge (Cooper) provides dedicated, named AI agents to professionals and businesses who want more than a chatbot — they want a capable, persistent digital chief of staff. Each client gets their own AI agent with a name they choose, a dedicated email address, Telegram access, and a private web chat portal. The agent learns their context, preferences, and workflows over time and handles real work: inbox triage, research, document drafts, briefings, project tracking, and more.

The business is built on a simple insight: people don't need more AI tools — they need an AI that actually knows them. ChatGPT and generic assistants reset every conversation. Our agents remember. They have context. They're configured around the client's life and work from day one.

Why the Economics Work

Costs are driven by Claude API usage, VPS server compute, and a minimal support burden. At average usage, a client on the Professional tier ($299/mo) costs roughly $28/mo to serve — leaving 91% gross margin. Even at heavy usage, margins stay above 79%. This is a high-margin SaaS business with extremely low variable costs.

What "Done Right" Looks Like

• A client sends a Telegram message and gets a genuinely useful response in under 2 minutes — not a canned reply
• The agent remembers context from weeks ago and applies it without being asked
• Onboarding takes 48 hours, not 2 weeks, and feels concierge-level
• Clients describe the agent to a friend as "weirdly good" — not "pretty good for AI"
• Zero data leaks between clients. Ever.
• When something breaks, we know before the client does

Year 1 Vision

We launch in concierge mode — 3 to 5 clients, all manually onboarded by John and Kent. We learn everything: what clients actually ask for, where friction is, what the agent gets wrong, what delights them. We use that learning to build the automated provisioning system. By month 3, we flip the switch to self-serve. By month 12, we have 50 clients generating $13K+ in monthly gross profit — and a system that can scale to 200+.

A Named Agent That's Yours

When a client signs up, they choose a name for their agent — Ryker, Maya, Scout, whatever feels right to them. That name becomes their agent's identity. The agent has a dedicated email address (ryker@agent.highside-re.com), a Telegram bot they can message any time, and a private web portal. From day one, the agent is configured around the client's context: their job, their priorities, their communication style, their current projects.

Three Interaction Channels

What the Agent Actually Does

Onboarding Experience

Day 1 — Welcome & Setup

• Client receives welcome email from John/Kent
• Short intake form: agent name, pronouns (optional), top 3 priorities, communication style preference
• Telegram bot activated, first message sent by agent introducing itself
• Web chat link delivered via magic link

Week 1 — Calibration

• Agent asks structured questions to build context (industry, key contacts, active projects, tools used)
• Daily briefings begin — client gives feedback on format/depth
• First real tasks completed — small wins to build trust
• Check-in call or message from John/Kent at day 5

Month 1 — In the Flow

• Agent has accumulated meaningful context about the client's world
• Client is using the agent habitually across channels
• 30-day review: what's working, what to improve, any tier changes

Daily Cooper Digest & New User Quick Start

Two core product features that drive early engagement, retention, and word-of-mouth.

🌅 Daily Cooper Digest

Every morning, the client's agent sends a structured daily digest email summarizing everything it did overnight, what's in motion, and 3 proactive insights it noticed. This is the single most powerful retention mechanic in the product — it makes the agent's invisible work visible and creates a daily habit loop. Clients who receive the digest don't churn. The digest includes:

• Completed Yesterday — specific tasks, emails handled, documents built, reminders set
• In Motion — active tracking items, monitoring, queued tasks
• 3 Things I Noticed — proactive ideas surfaced by the agent based on context (upcoming events, social signals, patterns)
• This Week So Far — metrics grid: tasks completed, emails processed, docs built, reminders set

The digest is sent from the agent's dedicated email address (e.g. cooper@agent.highside-re.com) and arrives before 7am at whatever time the client specifies. It's designed to feel like a message from a real executive assistant — personal, specific, and useful.

⚡ New User Quick Start List

On day one, every new client sees a curated Quick Start list organized into two categories: "Wow Me" (impressive one-shot deliverables) and "Make Me Unstoppable" (ongoing automations). This list exists to eliminate the blank-slate problem — new users don't know what to ask an AI agent, so we tell them exactly what to ask. Most clients pick one item, get a stunning result within the hour, and immediately understand the product's value. The Quick Start list is:

• Personalized by industry/role — real estate investors see deal analysis and market research; founders see investor prep; executives see board materials
• Evolving — over time, Cooper's suggestions reflect the client's actual usage patterns, not a generic template
• Action-oriented — every item has a clear "→ Ask Cooper" button that opens a conversation with the agent pre-seeded with context

These two features together — the Digest and the Quick Start list — form the product's onboarding and retention backbone. The Quick Start converts curious prospects into active users on day one. The Digest keeps them engaged for months.

What Makes This Different

Trust & Memory — Core Value Propositions

Privacy and persistent memory are not just technical features — they are the two most powerful selling points of the Cooper product, and they must be communicated clearly in every sales conversation. This section gives the full picture: how they work, why they matter, and how to talk about them.

Privacy Architecture — Why Client Data is Isolated

Every Cooper client runs in a completely isolated workspace on the server. This is not a shared SaaS platform. Here's what isolation means in practice:

• Separate workspace directory — each client has their own /data/.openclaw/workspace/ equivalent. No client can see or access another client's files.
• Separate agent process — each client's Cooper instance runs independently. One client's agent cannot query or interact with another's.
• Separate credentials & channels — each client has their own Telegram bot, their own email address, their own API key context. Cross-contamination is structurally impossible.
• No shared model training — client conversations do not feed into any training pipeline. What they tell Cooper stays in their workspace.
• Operator blind spot by design — John and Kent can verify that a client's agent is running and healthy. They cannot read the client's conversations, files, or memory. The privacy is not just policy — it's structural.

The sales framing: "Think of it as your assistant's private office. We gave you the key. We don't have a copy."

Memory System — The Technical Basis for Persistent Memory

Cooper's memory is built on a small set of plain-text files that live in the client's workspace. These files are loaded at the start of every session, giving Cooper full context about who the client is and what's happened:

• USER.md — the client profile. Name, role, company, family, communication preferences, active projects, key contacts. This is set up during onboarding and updated as Cooper learns more. It's how Cooper knows who you are before you say a word.
• MEMORY.md — curated long-term facts. Significant decisions, corrections, important context that must persist across sessions. Think of it as Cooper's notebook — the things it has decided are too important to lose.
• Daily notes (memory/YYYY-MM-DD.md) — a rolling log of each session. Key decisions made, corrections received, new contacts added, action items generated. Short, bullet-point format. Written by Cooper at the end of each session or significant interaction.

At the start of every session, Cooper reads USER.md, MEMORY.md, and today's (and sometimes yesterday's) daily note. This gives it a complete, current picture of the client's world — before the client says anything. That's the mechanism behind "Cooper already knows your context."

How Memory Improves Over Time — Daily Notes to MEMORY.md Distillation

Memory is not static. It grows and improves through a distillation process:

• Every significant interaction produces a daily note entry. Over time, these accumulate into a rich, granular history.
• Periodically (every few days), Cooper reviews recent daily notes and distills the most significant facts into MEMORY.md — pruning stale entries and elevating what matters.
• When a client corrects Cooper ("that's wrong, my partner's name is X"), Cooper updates MEMORY.md immediately and acknowledges it out loud. Corrections are permanent — they don't get lost between sessions.
• USER.md evolves too. As Cooper learns more about the client's world (new projects, new contacts, changes in priorities), it updates the profile file.

The result: a Cooper that genuinely gets smarter and more personal over time — not because the AI model changed, but because the context it works with becomes richer and more accurate with every interaction.

The Learning Curve — Day 1 Through Month 6

This is the narrative arc for sales conversations. Use it to help prospects visualize the long-term value of Cooper:

• Day 1 — Already useful. Cooper learns the client's name, role, work, and goals. Sets up morning briefing. Completes first Quick Start request. The client is already getting value before the end of their first day.
• Week 2 — Knows you. Cooper has accumulated enough daily notes to understand communication style, key people, and voice. It starts drafting emails that sound like the client. The client stops having to explain context — Cooper already has it.
• Month 1 — Gets you. Cooper knows what matters to this client specifically. It surfaces ideas proactively. It handles routine things without being asked. The client starts to notice they're getting less done themselves — because Cooper is handling it.
• Month 6 — Invaluable. Cooper knows this client's projects, history, and preferences deeply. Proactive suggestions are specific to their actual life. Drafts sound exactly like them. At this point, the client cannot imagine working without Cooper. Churn is essentially zero.

Messaging Guidance for Sales Conversations

When speaking with a prospect about privacy and memory, use these framings:

Overview

Highside AI Concierge (Cooper) runs on a single Hostinger KVM VPS (starting on KVM4, upgrading to KVM8 at 20+ clients). Each client is completely isolated — separate workspace directory, separate process, separate credentials, separate channels. OpenClaw is the agent runtime. Claude is the AI brain. Cloudflare handles DNS, routing, and the web chat frontend.

Multi-Client File System Design

Each client gets a fully isolated workspace on the VPS. Structure:

Email Architecture

Domain Setup

• Primary domain: agent.highside-re.com (Cloudflare DNS)
• Catch-all inbox configured at the domain level
• Each client gets a dedicated address: [agentname]@agent.highside-re.com
• Microsoft Graph API (already configured) handles send/receive

Email Routing Logic

Telegram Architecture

One bot per client — this is the cleanest isolation strategy. Each client has their own Telegram bot (e.g., @RykerAgentBot). No shared state, no cross-contamination, clear identity. Bots are created via BotFather during provisioning and stored per-client in credentials.

• Webhook registered to agent.highside-re.com/telegram/[client-id]
• Nginx reverse proxy routes to the correct OpenClaw session
• Each session is isolated — one session per client, always running

Web Chat Architecture

• URL pattern: agent.highside-re.com/chat/[client-id]
• Cloudflare Pages hosts the static frontend
• Magic link auth: client clicks link in email → JWT set → session authenticated
• WebSocket or SSE connects to the VPS for real-time responses
• No passwords stored — magic links expire in 24 hours

Server Sizing

Security & Data Isolation

• Filesystem permissions: Each client workspace is chmod 700, owned by a dedicated system user (agent-001, etc.). No cross-read possible at OS level.
• Credentials directory: Separate from workspace, never loaded into AI context. Keys are file-referenced, not embedded in prompts.
• No shared memory: Each OpenClaw session runs in isolation. MEMORY.md files are per-client and never cross-loaded.
• Daily backups: Automated encrypted backup of all client workspaces to an offsite location (Cloudflare R2 or B2). 30-day retention.
• Audit logs: Per-client session logs preserved for 90 days. Never surfaced to other clients.
• No AI training on client data: Anthropic enterprise API terms explicitly prohibit training on API data.

Phase 1: Manual Concierge Onboarding (Clients 1–5)

The first 5 clients are a learning lab. John and/or Kent handle every step personally. The goal is NOT efficiency — it's depth of understanding. Every friction point gets logged. Every "why can't it just..." gets written down. These notes become the spec for the automated system.

Step-by-Step: Manual Onboarding

1. Intake Form Sent (Day 0) — Client fills out: agent name, primary use cases, communication style, top 3 current priorities, timezone, email access preference.
2. John/Kent Reviews Intake (Day 0–1) — Manually builds USER.md, SOUL.md, and initial MEMORY.md for the client. Configures agent persona based on intake answers.
3. Infrastructure Provisioned Manually (Day 1):
   • Create workspace directory with correct permissions
   • Create BotFather Telegram bot, store token in credentials
   • Configure Microsoft Graph mailbox/alias for the agent email
   • Register Telegram webhook
   • Create web chat endpoint + magic link
   • Test all three channels end-to-end
4. Client Welcome (Day 1–2) — Personal email from John/Kent. Includes Telegram bot link, web chat link, and agent email address. Short video walkthrough (recorded once, reused).
5. Quick Start List Shown on Welcome Page (Day 1) — The client's welcome page includes the personalized Quick Start list, pre-configured for their industry and role. Two categories: "Wow Me" (one-shot impressive deliverables) and "Make Me Unstoppable" (ongoing automations). The list is the answer to "where do I even start?" — clients pick one item and have a result in their inbox within the hour.
6. Agent Introduction (Day 2) — The agent sends its first message via Telegram: "Hi, I'm [Name]. I've been set up to help you with [top priorities from intake]. Let's start simple — what's the one thing on your plate right now I can help with?"
7. Daily Digest Begins (Day 2 onward) — Each morning, the agent sends a structured Daily Digest email: tasks completed, what's in motion, 3 proactive insights, and a weekly stats grid. First 30 days include a "Did you know Cooper can..." tip at the bottom of each digest to surface capabilities clients haven't discovered yet.
8. Day 5 Check-In — John or Kent personally follows up: what's working, what's confusing, anything to fix.
9. Week 1+ Personalization — After the first week, the agent begins personalizing the Quick Start list and digest insights based on what the client has actually done — surfacing ideas that match their real usage patterns, not a generic template.
10. Day 30 Review — Formal check-in. Usage stats reviewed. Tier upgrade conversation if usage is high. NPS survey sent.

What We're Learning (Log Everything)

• Which parts of setup take longest → automate those first
• What questions clients ask most in first week → pre-answer in onboarding
• Where agent responses fall short → improve SOUL.md and skill config
• What clients rave about → double down
• What feels clunky about channels → UX improvements

Phase 2: Automated Provisioning (Clients 6+)

After 5 clients, we have a playbook. Now we automate it via Stripe webhooks + a provisioning script.

Internal Client Dashboard

A private, password-protected admin panel (hosted on Cloudflare Pages with Access protection) gives John and Kent full visibility into the business.

Client List View

• Client name, agent name, tier, MRR contribution
• Status: Active / Onboarding / Churned / Paused
• Last active timestamp across channels
• Days since onboarding, days until renewal
• Quick link to each client's workspace and logs

Usage Metrics (per client)

• Message count (by channel): email, Telegram, web chat
• Estimated Claude API token usage and cost
• Tasks completed, documents generated (rough count via log parsing)
• Usage trend: flat, growing, at-risk (declining)

Revenue Tracker

• MRR by tier, total MRR, projected ARR
• New MRR, churned MRR, net MRR change (monthly)
• Cost breakdown (server, API, tools) vs. revenue — gross margin live view
• Stripe sync: payment status, failed charges, upcoming renewals

Provisioning Tooling

A set of admin scripts (shell + Node.js) handle client lifecycle:

• provision-client.sh [client-id] [tier] — Full setup from scratch
• deprovision-client.sh [client-id] — Clean teardown + archive
• upgrade-tier.sh [client-id] [new-tier] — Adjusts any tier-based limits
• backup-client.sh [client-id] — Manual backup trigger
• status-check.sh — Verify all client sessions are running

Monitoring & Alerts

• UptimeRobot or Betterstack: Ping each client's Telegram webhook every 5 minutes. Alert on failure.
• Session health checks: Cron job every 15 minutes verifies all OpenClaw sessions are responsive. Auto-restart on failure, alert if restart fails.
• API error tracking: Claude API errors, Graph API failures → logged centrally, alerted on spike
• Disk + RAM alerts: VPS monitoring via Netdata or similar. Alert at 80% usage.
• All alerts go to: John + Kent via Telegram (the tool of the business)

Pricing Tiers

Trial: No Credit Card Required

Every plan starts with a 7-day free trial — no credit card required. After 7 days, the client chooses a plan and enters payment. Cancel any time during the trial with zero friction — one email to Cooper is all it takes.

Annual Billing

Offer 2 months free on annual commitment (pay 10 months, get 12). Present at onboarding and again at month 3.

• Personal Annual: $1,990/yr (effectively $166/mo)
• Professional Annual: $3,490/yr (effectively $291/mo)
• Executive Annual: $5,990/yr (effectively $499/mo)

Flexible Code System

John and Kent can generate custom discount codes on demand via Stripe. Each code can be configured for:

• Any discount percentage — 10%, 20%, 30%, whatever makes sense
• Any duration — one month, 6 months, or permanent (forever)
• Single-use or multi-use — made for one person, or reusable by many
• Extended trial length — e.g. 14 days instead of 7
• Restricted to a specific email address — Stripe supports this natively

Use cases:

• Sending a personal friend a code for 2 free weeks + 30% off forever
• A business contact gets 15% off for 6 months
• A referral partner gets a reusable code for their network

No predefined tiers or code names needed — full flexibility on demand. Codes are entered at checkout on the Stripe payment form.

Referral Program

After a client's first week, Cooper proactively asks if they know someone who could use this. Each client gets a personal referral link. When their referral signs up, the referrer gets one free month. No limit on referrals.

Stripe Setup

• Products & Prices: Create three subscription products in Stripe (Personal, Professional, Executive) with monthly + annual price variants
• Webhooks: customer.subscription.created → trigger provisioning; customer.subscription.deleted → trigger deprovisioning; invoice.payment_failed → alert + grace period email
• Billing portal: Enable Stripe Customer Portal for self-serve plan changes and payment updates
• Tax: Enable Stripe Tax for automatic sales tax calculation (TX has software service exemptions — verify with accountant)
• Trials: 7-day free trial on all plans, no card required at signup
• Discount codes: Create via Stripe Promotions — configure on demand per the Flexible Code System above

Usage & Overage Nudges

• Usage tracked per client per billing cycle
• Heavy usage clients nudged toward Executive tier proactively by their agent
• Executive tier: designed for power users — API cost spikes trigger internal alerts for ops review

What This Means Operationally

The Monitoring-First Principle

Before launching anything — feature, client, server upgrade — the monitoring must be in place first. You don't add monitoring after something breaks. You need to know it broke before the client does. This means:

• UptimeRobot alerts configured before any client goes live
• Session health check cron running before first client provisioned
• Log review cadence established (daily for first 30 days)
• Disk and RAM alerts configured on VPS
• All alerts go to both John and Kent via Telegram

Growth Targets

Quality Metrics

Year 1 Revenue Model

Operational Metrics to Track Weekly

• New clients this week — target 2–3/week by Month 3
• Churned clients this week — flag any churn immediately, understand why
• Support tickets / issues raised — rising = something's wrong with product or agent quality
• Uptime % across all client sessions — target 99.5%+
• Average response time (Telegram) — target <90 seconds during business hours
• API cost per client (Claude) — watch for anomalies indicating runaway sessions
• Upgrade rate — % of Starter clients upgrading to Pro within 90 days (target: 20%+)