Tonight's Agenda

01 Lunatechs Intro Hanley

02 OpenClaw — Another Look Hanley

03 Unleashing the Full Potential of OpenClaw LokLok

04 LUCKY DROP!

05 OPEN MIC

Hanley Leung

Build Cool Stuff

Hello

I'm Hanley

Hometown

San Francisco

Currently

San Francisco ↔ Hong Kong ↔ Shanghai

US Citizen · HK Permanent Resident · 12 Years in Shanghai

Education

UC Berkeley

Loves

Tech, Travel, Fitness, Gaming

Apple

Apple Vision Pro

EA

Pogo, FIFA Online 2, NBA Street Online

Yahoo!

Yahoo! Mail, HotJobs, Classifieds

R2 Digital

make Froyo!, Teddy Bear Workshop, Mochi Monster Blast

TAISO AI

AI Agentic Platform
Natural Language
Business Automations

Kids Challenging Kids
AI From Line 1
UGC Hypercasual Game Platform

Hanley Talks Tech

Tech Reviews & Tutorials

▶

Hanley Talks AI

AI News & Insights

▶

Your tribe of Technology Lunatics.

CharterBuild. Learn. Share. Connect.

MissionChange the world.

EventsHikes, Meetups, Office Hours, Hackathons, Industry Excursions

LinkedIn Group

Join the conversation

WhatsApp Community

Get involved today

We're Looking For

Speakers

Anything cool, even non-tech

Sponsors

Would love to be able to feed everyone

Organizers

Two more people to help organize events & community

Chapters

Already have a Silicon Valley chapter organizing

LinkedIn Group

WhatsApp Community

The AI Agent That Broke the Internet

200,000

GitHub Stars

One developer. One weekend. One idea.

One of the fastest-growing open-source projects ever.

Your New Coworker

Lives in your chat apps
Runs on your machine
Works while you sleep
Acts — files, email, browser, code, calendar

It Works For You

Your Message

→

AI Thinks

→ loop ←

Agent Acts

→

Result

The loop repeats until the task is done — or the AI gives up.

It Works On Its Own

⏰

Cron Jobs

Scheduled tasks on any cadence — reports, backups, morning briefings, batch jobs.

"Every Monday 9am, summarise last week's PRs"

💚

Heartbeat

Periodic check-ins — monitors systems, stays aware, acts when something changes.

"Every 30 min, check if staging is still healthy"

⚡

Webhooks

Reacts to external events instantly — emails, messages, alerts, deployments.

"When a new email arrives, triage and reply"

No human in the loop — it just handles it.

It can even set up its own scheduled tasks.

Under the Hood

External AI

AI Brain (LLM)

Anthropic · OpenAI · Google · Ollama

Cloud AI: Anthropic / OpenAI / Google Local AI: Ollama (free, private)

OpenClaw

Gateway

router + auth + sessions

Web UI

dashboard + config

Memory

files + vector search

Skills

markdown manuals

Cron

7x24 automation

Chat Connections

WhatsApp · Telegram · Slack · Discord

Tools

Full Computer

shell, browser, files

Internet APIs

web, email, services

Magic Moments

👋

I said "Hi" to my Mac Studio on WhatsApp — and it replied

📊

I asked it to scan websites, process the data, and generate a visual report — it worked on the first try

🔔

A message appeared from the AI that I didn't ask for — it just ran on schedule

🧠

It referenced a conversation we had days ago — without being reminded

Why Did It Blow Up?

Runs as You

Full access to your computer and everything on it. Files, apps, APIs, the internet.

Agents Just Go

Give it a task, walk away. It figures out the steps, picks the tools, and executes.

Triggered & Scheduled

Cron jobs, heartbeats, webhooks — it works while you sleep.

Chat From Your Phone

Message it on WhatsApp or Telegram. Go enjoy your life — it keeps working.

What People Use It For

Email triage
Shopping & negotiation
Coding & DevOps
Content pipelines
Trading & finance
Research & reports

CEO Content Pipeline

Fu Sheng (Cheetah Mobile) spent Chinese New Year building a multi-agent system. Ran a full content pipeline:

6 articles · 1M+ views · 300K video views — while on vacation.

AI Bought a Car

Datadog engineer told OpenClaw to buy a Hyundai Palisade. It emailed dealers, negotiated, played them against each other.

Saved $4,200 below sticker — no human phone calls.

Polymarket Trading Bot

An OpenClaw agent automated prediction market trades on Polymarket, exploiting timing inefficiencies at scale.

Claimed ~$150K profit — went massively viral.

The Big Questions

Is it safe? Honestly?

It can do anything your user account can do on your computer
It can crawl your machine and find passwords & API keys
Any API it can access, it can call — or call wrong by accident
It can be tricked into leaking credentials, compromising accounts
It can change your security settings
If you allow Python — full access to everything via scripts it writes

Same risks as giving a stranger your unlocked laptop. Start small, build trust.

How much?

OpenClaw itself — free & open source
You pay for hosting + an AI brain
I pay $200/mo for Claude MAX
I've burned $100/hr on GPT on one project
It adds up fast

Need to code?

Right now? Yes, honestly
Lots of terminal & JSON editing
Hard to debug
Not plug-and-play yet

Is my data safe?

Local model = AI companies can't train on your data
But doesn't fix the security risks above
The agent still has full access to your machine

Can it mess up?

HELL yeah — but so do you, right?
Different mistakes, not more
It just makes them a hell of a lot faster than you

Cost runaway?

AI provider — set a limit, easy
Other APIs it touches? Can go crazy
Put cost controls on everything it has access to

Do I need a Mac Mini?

No — any computer works
Local models: M4 Pro 64GB is nice but slow
Just use a cloud API
That said — Macs are niiiice

Overhyped — or the Next Big Thing?

It has a real chance to become the Apache of AI agents — the defacto open-source standard.

Hundreds of contributors are shipping code every week.

Apple, Microsoft, and Google will build their own — tied to their ecosystems.

I'm building an Apple ecosystem version. Architecturally very different.

Not whether you'll have an AI agent.

When. And whose.

Apple

iOS + macOS + Private Cloud Compute

Microsoft

Microsoft 365 + Windows

Google

Google + YouTube + Search

Huawei

HarmonyOS

Tencent

WeChat

Alibaba

Taobao + Alipay

Try It Tonight

A Dedicated Computer

Do NOT install on your personal computer
Work computer? Only if you don't mind getting fired if it goes sideways — or your boss said go for it
Best: a cheap dedicated Mac Mini or old laptop
Install OpenClaw + an AI API key

Cost: ~$600 USD / ~$4,700 HKD Mac Mini or old laptop
+ AI API $5–200 USD / $40–1,600 HKD per month

Cloud Server

One-click deploy solutions already exist
Runs 24/7 — no hardware to manage
Sandboxed — can't mess up your personal stuff
Be careful of contamination — don't link personal accounts or data
Connect via WhatsApp, Telegram, or web
Best for always-on cron jobs & heartbeats

Cost: $5–50 USD / $40–400 HKD per month server
+ AI API $5–200 USD / $40–1,600 HKD per month

▶

Now Go Build.

Change the World.

npm install -g openclaw@latest

docs.openclaw.ai/install

5,700+ Skills

What Are Skills?

Plain markdown files that teach the AI how to do specific tasks. Anyone can write one. They get injected into the system prompt when triggered.

# Grocery Shopping
## Trigger
"buy groceries" or "shopping list"
## Steps
1. Check the shared list
2. Find best prices online
3. Place the order
4. Confirm with user

Not a business moat — too easily copied or recreated. Great for getting AI working for you, but don't build a company on skills alone.

Marketplaces

ClawHub

OpenClaw's own skill marketplace. Community-contributed, moderated.

GPT Store / Gems

OpenAI & Google have similar concepts — custom instructions for their AIs.

MCP Servers

Anthropic's Model Context Protocol — tools as plug-in servers.

Examples

github-pr-review 1password discord-mod web-scraper email-digest calendar-sync stock-alerts home-assistant expense-tracker meeting-notes

Tools

Built-in Tools (76)

Tools are how the AI actually does things. Each tool is a function the AI can call — read files, send messages, search the web, run code.

message_send web_fetch web_search exec_command memory_search file_read file_write browser_navigate cron_schedule image_generate canvas_draw + 65 more...

Current State

Basic implementations work

Tools today are functional but rough. They get the job done, but they're not optimized for efficiency, security, or privacy.

Where It's Going

→ Efficiency — smarter caching, fewer API calls
→ Capability — richer browser, vision, voice
→ Privacy — sandboxed execution, data isolation
→ Security — per-tool permissions, audit trails

Every tool will evolve a lot. Massive innovation ahead.

Who Built What

Frameworks OpenClaw Relies On

Pi

by Mario Zechner

The agentic loop core
Claude / GPT / Gemini / Ollama

Baileys

WhatsApp Web API

Reverse-engineered protocol
Community maintained

grammY / Discord.js / Bolt

Chat platform SDKs

Telegram, Discord, Slack
iMessage, Mattermost

What OpenClaw Built

700-line system prompt 76 tools Memory system Security pipeline Cron & scheduling Gateway & routing Skills system Device pairing Web UI Plugin SDK

Zero new AI.

Brilliant integration of existing open-source pieces.

How the System Prompt Is Built

Every message, OpenClaw dynamically assembles the prompt:

1

Base System Prompt

Identity, rules, personality, safety boundaries

2

Dynamic Tool + Skill Selection

Not fixed — AI has access to ALL tools & skills, system picks relevant ones per message

3

Memory Recall

Vector + keyword search pulls relevant past context into the prompt

4

Conversation History

Recent messages + previous tool call results

5

Your Message

Appended at the end — the AI reads everything above before responding

Old Way vs New Way

Old: Fixed worker with fixed skills. 76 tool definitions + all skills crammed into every prompt.

New: Dynamic assembly. Agent has access to everything, but only the relevant tools and skills are injected moment-by-moment.

Like a worker with a whole toolshed — grabs what's needed for each task instead of carrying every tool at once.

The Trade-off Remains

Even with dynamic selection, more context = less room to think.

The smarter the selection, the smarter the AI stays.

This is where the real engineering challenge is.

Memory System

Hybrid Search

Vector Similarity + BM25 Keyword

File Store

MEMORY.md + transcripts

SQLite Index

Vector embeddings + FTS

Embedding Providers

OpenAI → Gemini → Voyage → Local
Auto-fallback chain

Smart Ranking

Temporal decay weighting
MMR deduplication
Query expansion

74 Source Files

Incremental sync + file watching
Session-level context windows

Software Components

OpenClaw is a self-hosted gateway

A single process on your hardware that connects your chat apps to AI agents — no hosted services required.

🧠

Agent Engine

AI loop + tools

🌐

Gateway

Router + auth + sessions

💬

6 Channels

WhatsApp · Telegram · Slack · Discord · iMessage · Mattermost

🧩

Memory

Vector + FTS

🛡️

Security

Injection + SSRF protection

⏰

Cron & Jobs

Scheduling + heartbeats

📜

Skills

55+ bundled · Markdown

🖥️

CLI + Web UI

Config + dashboard

The Context Window Problem

The Problem

→ AI has a fixed-size "working memory"
→ Long conversations overflow the window
→ AI literally forgets what you said earlier
→ Costs scale with context size ($$$)

The Ability Paradox

→ Every tool definition eats context space
→ Every skill injected eats context space
→ More abilities = less room for thinking
→ Add more capability, AI gets dumber

How OpenClaw Handles It

→ Memory system stores long-term knowledge
→ Vector search retrieves relevant context
→ Session transcripts for conversation history
→ Smart summarization before overflow

AI Context Window — ~200K tokens

System Prompt ~5K

Tool + Skill Definitions ~20-50K

Memory Recall ~5-20K

Tool Results ~10-50K

Actual Thinking Space whatever's left

OVERFLOW → forget

Real Numbers

200K tokens ≈ 150K words ≈ 500 pages
76 tools can eat 20-50K just in definitions
Add skills on top → even less room to think
A busy session burns 50-100K tokens
More ability = less intelligence

Selina Li

Co-founder · HK tech star

Bubble Lab

Y Combinator Backed

→ More production-ready than OpenClaw right now
→ Set up automations with natural language
→ Worth trying if you want something that just works today

Mention "Lunatechs / Hanley" — sending Hong Kong support!

bubblelab.ai