Use ↑ ↓ / PgUp PgDn / click dots
Founder meeting · 17 May 2026

ClawOne: the private AI box for working agents

Tomorrow is not about a polished pitch. It is about founder alignment: who we start with, what the first box must do, and what we refuse to overbuild.

Singapore firstHermes runtimeAgent workflows12-month refresh
Singapore skyline at night
Singapore is the test market: compact, high-trust, and dense with commission-driven agents.
01
Thesis

Do not sell a generic AI box. Sell fewer missed follow-ups, faster replies, and memory that does not leak out of the business.

Private box
→
Daily workflows
→
Client memory
→
Country playbooks
→
Better margins

The first product should feel boringly useful. The bigger company can come later.

02
Positioning

“ChatGPT is the brain. Hermes is the memory and workflow manager. ClawOne is the private box that keeps it running for your business.”

Trusted appliance

Router-style managed device, not infrastructure rental.

Workflow layer

Replies, follow-ups, owner updates, reminders.

Memory layer

Client history, promises, playbook learning.

03
Architecture decision

ClawOne runs Hermes. Hermes absorbs only the features we need.

Ship

  • Hermes as product runtime
  • Telegram / WhatsApp interface
  • Workflow coach + reminders
  • Client memory + playbook learning
  • Model routing

Borrow, don’t ship separately

  • OpenClaw: tools, scheduling, health checks, routing
  • Nanoclaw: ingest, fact extraction, memory graph/search
  • No three-layer support burden
  • No duplicated memory/runtime state
04
Why agents first

Agents already live inside messy conversations. That is where ClawOne can earn its keep.

Commission-driven

If the box helps close or retain one deal, the math is easy.

Messenger-native

The job happens in Telegram, WhatsApp, calls, images, and half-written notes.

Memory-heavy

Clients, owners, objections, timelines, promises, viewing notes.

Admin-heavy

Updates, replies, reminders, briefs, and repetitive market explanations.

Real estate meeting
Use visual language that feels close to sales work, not abstract AI infrastructure.
05
Product boundary

Internal-only assistant. Human approves and sends.

What ClawOne does

  • Triages shared messages
  • Drafts replies
  • Remembers promises
  • Prepares owner updates
  • Suggests next actions

What ClawOne does not do first

  • No outward-facing chatbot
  • No unsupervised customer replies
  • No formal valuation claims
  • No fragile data-source promises
06
MVP v0

First box should prove deployment reliability before workflow breadth.

1

Onboarding

Browser setup, WiFi, device name, status.

2

Model setup

Preferred Codex / cloud model test.

3

Messaging

Bind Telegram, send test, status commands.

4

Hermes

Start memory, workflows, healthcheck.

07
Workflow wedge

Start with practical agent jobs, not futuristic autonomy.

Lead / reply copilot

Turn shared chat context into approved reply drafts.

Owner update

Client-friendly property/market update with caveats.

Viewing brief

Summary, talking points, risks, questions to ask.

Follow-up memory

Remember the next action and draft the nudge later.

Market pulse

Weekly area/project summary with source labels.

Playbook learning

What worked, what failed, what to avoid.

08
Phase 1 model

Keep inference risk off our books until the workflow is proven.

S$59–69

Working monthly ClawOne fee for the appliance, workflows, onboarding, support, and updates.

S$199–299

Deposit range to discuss. It must deter non-return without killing adoption.

12 mo

Contract term and renewal-refresh direction.

Fact-check: Codex is currently included in ChatGPT plans, but packaging can change. Treat BYO Codex as a Phase 1 simplifier, not a permanent dependency.

09
Hardware lifecycle

12-month refresh turns fast-moving AI hardware into a renewal loop.

Tier 1 new device
→
12-month contract
→
Renewal refresh
→
Wipe + refurbish
→
Tier 2 trials

Why 12 months

AI hardware and local models move quickly. Refresh keeps the box current and gives customers a concrete reason to renew.

Inventory risk becomes option value

Returned units become lower-capex stock for Malaysia, Indonesia, Philippines, and Vietnam pilots.

10
On-prem constraints

Optimize for reliability + memory before local AI ambition.

  1. Onboarding, Telegram, reboot recovery, `/healthcheck`.
  2. Structured local memory/search before generation.
  3. Workflow commands over autonomous always-on agents.
  4. Cloud burst for complex drafting and reasoning.
  5. Small local models only for tags, reminders, short summaries.
  6. Nanoclaw-style graph/wiki runs scheduled or on-demand.
11
Phase 2

Local model box captures power users without promising “unlimited AI”.

Best users

  • Heavy daily users
  • Teams/agencies
  • Privacy-sensitive workflows
  • Users hitting cloud limits

Pricing hypothesis

  • S$99 only if hardware is separate
  • S$149–299 all-in if hardware is bundled
  • Say: no per-token cloud bill for supported local workflows
12
Phase 3

Margin expansion through model orchestration.

Internal / investor logic

  • Route simple work to local/open models
  • Use premium models only when needed
  • Cache/templates for repeated work
  • Evaluators reduce rework

Customer language

  • Better outcomes
  • Reliable workflows
  • Local compliance/language support
  • Automatic model choice

Do not sell “token arbitrage” to customers. Sell reliability and workflow quality.

13
Singapore market

We have one hard number. The rest should stay labelled as a working model.

36,816

CEA-reported property agents in Singapore as at 1 Jan 2026.

63K–73K

Working total agent TAM after adding an unverified insurance/FA estimate.

S$52M–60M

ARR model at S$69/month. Correct arithmetic; underlying total-agent range still needs validation.

Sources: CEA industry statistics; data.gov.sg CEA Salesperson Information; MAS representative register for source category, not count.

14
Regional roadmap TAM

The regional story is promising, but not yet investor-grade.

1.06M–1.71M

Directional roadmap agent TAM across SG/MY/HK/PH/ID/VN/AU.

S$878M–1.42B

ARR model at S$69/month. Useful for scale intuition, not final diligence.

50K–100K

7-year operating milestone: S$41M–83M ARR before upsells.

Founder takeaway: Singapore proves the wedge. SEA creates the scale. Phase 2/3 improves gross margin.

15
Expansion sequence

Proof market → adjacent localization → partner-led scale.

0

Singapore property

Prove onboarding, Telegram, BYO Codex, ROI.

1

Singapore depth

Add insurance/FA workflows after usage proof.

2

Malaysia

Close adjacency, multilingual overlap, WhatsApp workflows.

3

HK + SEA

HK high ARPU; PH/ID/VN via local partners.

4

Australia

Higher ARPU, more competition, later entry.

16
Defensibility

A static pack is copyable. A living, licensed, customer-adapted workflow system is harder.

Country-specific workflow packs
Customer memory + playbook learning
Signed/licensed updates
Private evaluators + routing logic
Source-aware QA
Distribution into agent communities
17
Founder decisions

What we need to align on tomorrow.

Policy

  • Confirm property → insurance/FA sequence
  • Confirm S$59–69 + deposit offer
  • Set deposit, return, damage terms
  • Define refurb QA and inventory limit

Build

  • Confirm Hermes-only runtime
  • Pick first hardware BOM
  • Assign QA/OTA/support ownership
  • Select first 5–10 pilot users
18
30 / 60 / 90

Execution path after alignment.

30 days

Finalize pilot offer, contract/deposit/refresh terms, MVP v0 scope, first hardware build, pilot list.

60 days

Build pilot image, test onboarding/model/Telegram/Hermes, QA recovery, prepare reply-copilot demo.

90 days

Deploy pilots, track usage/support/ROI, decide property expansion vs insurance/FA, validate lifecycle economics.

19
Fact-check and photo notes

What is verified, what is still a founder assumption.

Verified enough for internal use

  • CEA reports 36,816 property agents as at 1 Jan 2026.
  • data.gov.sg hosts official CEA salesperson data.
  • Codex is currently bundled into ChatGPT plan tiers.
  • Revenue figures are arithmetic from stated assumptions.

Keep caveated

  • Insurance/FA count needs stronger extraction.
  • Regional country counts are directional.
  • Deposit, refresh, QA, warranty, and inventory limits are policy decisions.
  • Phase 2/3 economics still need hardware/support modelling.

Photos: Wikimedia Commons Singapore skyline and real-estate office imagery. Detailed notes: fact-check-founder-meeting-2026-05-17.md.

20
Closing thesis

Start narrow. Prove daily utility. Use hardware lifecycle for SEA access. Capture margin through orchestration.

Appliance wedge. Workflow moat. Regional scale. Improving inference margins.

21