OpenClaw
Live Tonight · A Talk by Eric

OpenClaw

The iPhone Moment for AI — and how to wield it without setting your life on fire.

Released Jan 30 2026 · 250K+ stars · fastest-growing OSS, ever

Chapter ISetting Context

Before I begin — hold questions until the end. Most will get answered.

The Launch

OpenClaw was released January 30, 2026 — and went viral.

The fastest-growing open-source project ever to exist. 250K+ GitHub stars by early March.

250K+
GitHub stars
~2 days
Release cadence
2026.5.18
Stable channel
Jensen says

The CEO of NVIDIA called it the “iPhone Moment” for AI.

And he's not wrong.

The Shift

For those who spend time building with ChatGPT, Gemini, Claude — once you start using OpenClaw you'll recognize how it's leveling up your productivity.

The Key Shift

Instead of asking an AI questions, you delegate tasks.

OpenClaw can read messages, send emails, manage calendars, automate workflows, and act across real systems — all through messaging apps you already use.

WhatsAppDiscordSlackTelegramiMessage
HOWEVER
Warning

OpenClaw is a very permissive system.

The Plan

I'm going to cover my workflow AND how I keep it secure.

Chapter IIDisclosures

AI moves miles each day.

Anyone that says they are an expert or on the cutting edge probably isn't. You can spend all your time researching the latest and greatest — or you can build.

ChatGPTClaudeGeminiCursorCopilotDevinReplitv0BoltLovableWindsurfCodyPerplexityGrokMistralLlamaOllamaLangChainCrewAIAutoGPTOpenClawDispatchSkillsMCPChatGPTClaudeGeminiCursorCopilotDevinReplitv0BoltLovableWindsurfCodyPerplexityGrokMistralLlamaOllamaLangChainCrewAIAutoGPTOpenClawDispatchSkillsMCP
01 · Optimize your flow

Optimize the flow that works best for you

Focus on what gives you the leverage you need to get to your end goal. Personally, I lean heavily on Claude because I understand their ecosystem and platform the most.

What I'm demoing today is what I use in my workflow. I'm sure you all have your own tools — I'm not here to say mine is the golden standard. It simply works best for me.

02 · Pattern match

Pattern match across tools

The meta-takeaway: across tools, the most important thing you can do is pattern match. Almost every tool, model, or platform has the same concepts, just applied slightly differently.

OpenClaw's core loop — input → context → model → tools → repeat → reply — is the same pattern Claude Code uses. Every serious agent framework runs some version of it.

03 · Ask the model

When in doubt, ask

When you're stuck, just ask the model how you can work with it better or what you're not fully utilizing.

The model is not a vending machine — it's a collaborator. Treat it like one and your results compound.

AI is an infinitely patient explaining machine.
Peter Steinberger, OpenClaw's creator
Chapter IIIStrategy

Follow the crowd, selectively.

I wasn't an early adopter of OpenClaw. Same security concerns as everyone else — and I waited until I could isolate it without buying a brand new Mac Mini.

Sage Advice

If you wait around long enough, someone will build the tools you need to do your job — meaning you don't have to.

Early-career advice that aged well

01Stable

Got a more stable version of the software (the project releases roughly every 2 days — the early versions were rough).

02Crowdsourced debugging

Got a lot of people smarter than myself to figure out the problems with using it.

03Best practices

Community-contributed documentation, security guides, and best practices emerged.

Chapter IVWorkflow Evolution

My personal journey, in three phases.

01Phase

Prompting

Same as using ChatGPT or Claude, just in a different wrapper.

You type. It answers. The agent is a chat surface — pure request/response, you driving every interaction.

02Phase

Co-piloting

Running tasks side-by-side, checking in frequently, holding its hand.

You start a task. It takes a swing. You correct course. Trust is small but growing — you're together at the wheel.

03Phase

Delegating

"I'm going to bed, you make the decisions."

You stop watching. Standing orders, heartbeats, hooks — the agent runs the loop overnight and reports in the morning.

From request/response → to colleague → to autonomous teammate
Chapter VSetup

Demystifying the setup.

This is probably the most painful part — but we're not in February anymore. Things have gotten a lot easier.

Getting started

Go to openclaw.ai and follow the Quick Start — the goal is to start using it, you don't have to have it completely tricked out on Day 1.

HOWEVER — OpenClaw is very permissive by default. Assume it can execute commands, access files, and automate workflows on your machine.

My actual setup process

OpenClaw has docs on this, but I just used ChatGPT to give me step-by-step instructions from a fresh VM.

The pain is front-loaded — once you're set up, you're set up.

zsh
$ curl -fsSL https://openclaw.ai/install.sh | sh
$ openclaw init --name clank
$ openclaw channel add discord --token $DISCORD_TOKEN
$ openclaw start
   ↳ gateway listening on :7421
   ↳ heartbeat scheduled (5m)
Hosting Options

My recommendation: run it in a VM on your computer

  • Isolates OpenClaw from your personal data — this is the key security boundary.
  • You don't need a $700 Mac Mini.

Or: buy a hosting plan

Hostinger, Viktor, etc.

  • Easier to get running, but you don't own the machine — that has its own risks.
  • Token spending constraints with subscription plans (at least with Anthropic's $200/mo plan).
Chapter VIData Segregation & Security

The concept of zero-trust.

Safest approach: separate accounts for everything, share only what's needed. Pain to set up — but once you do it, you're done.

Agent ID · v2026.5online

Meet

Clank

Eric's OpenClaw agent

Owned accounts

  • iCloud
  • Discord
  • Email
  • 1Password
  • GitHub
  • Claude (shared)
SCOPE · isolatedPOSTURE · zero-trust

Why all of the above? Because of Integrations.

OpenClaw's power comes from connecting to external services, and every connection is a potential attack surface.

ClawHubPublic skills registry (2,800+)
⚠ buyer beware

Contains Skills and Plugins that extend what your agent can do. These are NOT endorsed by OpenClaw — buyer beware. Always do your own diligence before downloading. You reduce attack surface by following zero-trust and cordoning OpenClaw off from your personal data.

MCPsModel Context Protocol

MCP establishes three core primitives: Tools (functions the agent can call), Resources (data sources it can read), and Prompts (reusable workflow templates). The key point: MCP lets OpenClaw dynamically discover and connect to any compatible service.

APIsDirect service integrations

Direct integrations with external services. Voice via ElevenLabs and Whisper, browser automation, and more. Sky is the limit — and so is the attack surface.

Chapter VIIKnowledge Base

Sharing data — selectively.

Once your agent is isolated and secured, you selectively share what it needs.

Shared cloud drives

Google Drive, Dropbox

GitHub repos

Code & issues

Obsidian vaults

Personal knowledge

Project directories

Scoped local access

The principle

Your agent gets access to what it needs, not access to everything.

Chapter VIIICore Concepts

OpenClaw, under the hood.

Now that you have the setup and security foundations, here's what's actually running.

VIII · A

The Three-Layer Architecture

Think of OpenClaw like a body. Hover or tap a layer to inspect it.

Layer 01
Channel Layer
Ears & Mouth
Layer 02
Brain Layer
Agent Runtime — the LLM
Layer 03
Body Layer
Hands — tools, browser, files, memory

Channel Layer

Ears & Mouth

Messaging adapters. WhatsApp, Telegram, Slack, Discord, iMessage — all connect to one Gateway process. These normalize incoming messages from any platform into a standard format.

Brain Layer

Agent Runtime — the LLM

Where the LLM reasoning happens. Your agent's instructions, personality, model config, and memory live here. Model-agnostic: Claude, GPT, Gemini, local models via Ollama — you pick the model, OpenClaw handles routing.

Body Layer

Hands — tools, browser, files, memory

Tools, browser automation, file access, and long-term memory. This turns conversation into action: opening web pages, filling forms, reading documents, sending messages on your behalf.

The Gateway

The hub that ties it all together — a long-running process that manages sessions, routing, and dispatching.

VIII · B

Agents

Your AI persona, configurable in ~/.openclaw/openclaw.json.

  • · Has a name, role description, and model assignment (primary + fallback chain).
  • · Can run one agent (default) or multiple specialized agents side-by-side (e.g., email triage, scheduling, research).
  • · The agent's identity, rules, and capabilities are composed from markdown files on disk — not code. This is what makes it accessible.
VIII · C

Tools vs. Plugins vs. Skills

People confuse these — they're three distinct layers. Tap any card to flip it.

VIII · D

Memory

LLMs are stateless — every conversation starts fresh. OpenClaw solves this.

You can read your memory files in VS Code while the agent uses optimized search under the hood. Memory survives restarts, updates, and migrations.

01

MEMORY.md

Long-term facts in human-readable markdown.

02

Daily logs

Running context as the day unfolds.

03

SQLite + vectors

Machine-efficient search layer underneath.

VIII · E

Automation: Cron, Heartbeat, Hooks

This is what makes OpenClaw autonomous — the shift from reactive to proactive.

Cron Jobs

Your agent's calendar

Precise scheduling.

  • · “Every Friday at 5 PM, generate a weekly report”
  • · “Every morning at 8, check my email and surface what's urgent”

Persisted in ~/.openclaw/cron/jobs.json— survives Gateway restarts.

Heartbeat

Your agent's pulse

The subtlest and arguably most powerful feature.

Wakes up every N minutes, runs through a checklist (HEARTBEAT.md), decides if anything needs attention, stays quiet if nothing matters.

Feels like a colleague quietly keeping things on track.

Hooks

Event-driven triggers

Something happens — GitHub push, new email via webhook — and the agent reacts.

Paired with Standing Orders: persistent context and authority boundaries telling the agent what it's allowed to do.

Key Insight

“OpenClaw won't do anything autonomously until you configure the triggers.” The power is dormant until you set it up.

Cron for precision, Heartbeat for awareness, Hooks for reactions.

VIII · F

The Seven-Stage Agentic Loop (for the curious)

Every message flows through these seven stages.

1
Normalize
Standardize input from whatever platform.
2
Route
Which agent handles this?
3
Assemble Context
Package history, memory, instructions, skills.
4
Infer
Send to the LLM.
5
ReAct
Reason, call tools, integrate results.
6
Load Skills
Pull in capabilities on demand.
7
Persist Memory
Save what matters for next time.

This is the same pattern underlying Claude Code, LangChain, CrewAI, and every serious agent system. OpenClaw just makes it tangible, file-based, and readable. This is the pattern matching I was talking about earlier.

Chapter IXStatus Check

Now What?

You now have an always-on, fully autonomous agent.

Sky is the limit.

Chapter XHelpful Tips

A little janky, a lot more powerful.

OpenClaw can be a little more janky than using ChatGPT or Claude — but it's also much more powerful.

clank@vm · ~/.openclaw
$openclaw tips --next
tip 01Observe the behavior

01
Observe the behavior
Watch when it trips over itself. Sometimes it takes a while to respond — that means it's thinking.
02
Prod it occasionally
There's a timeout issue I haven't fully fixed yet. The heartbeat configuration helps here.
03
Software is imperfect
General rule, by the way — doesn't mean it doesn't have utility.
04
Figure out your flow
Find what works for you and iterate.
05
Ask the AI for feedback
Ask it when you're stuck — or why IT'S stuck. Remember: infinitely patient explaining machine.
06
Stay a couple versions behind
Releases every ~2 days across alpha / beta / stable. The stable channel is your friend. Current stable: 2026.5.18.
Chapter XILive Demo

Live demo & workflow examples.

A taste of what delegation actually looks like. (Switching to the Discord channel now…)

#clank
Discord · DM
live
You
Clank, summarize my unread email and book lunch with Priya next week.
Clank
On it. Scanning inbox… 23 unread, 4 actionable.
gmail.listskills.email_triage
Clank
Priya is free Tue 12:30 or Thu 1:00. Picked Thu 1:00 — closer to her free block. Reply sent. Calendar held.
calendar.findcalendar.creategmail.send
Clank
Daily summary persisted to MEMORY.md. Anything else?
memory.persist
Clank is thinking…
Demo Notes
  • · Channel: Discord DM with Clank
  • · Skills loaded: email_triage, calendar_assist, daily_summary
  • · Standing Orders: don't send mail to anyone outside whitelist without preview
  • · Heartbeat: 5 minutes — surfaces anything urgent on its own
If the demo gods are kind

We'll do a live cron-triggered weekly report, a heartbeat-driven email surface, and a hook reacting to a GitHub push — all in this terminal window.

If they're not, we'll move on. Software is imperfect.

Chapter XIIMy Two Cents

Why learn all of this?

Wake-up moment

OpenClaw was a wake-up moment for the frontier models.

The concepts that OpenClaw introduced — persistent agents, heartbeat, skill registries, multi-channel delegation — while not the most stable, are powerful tools that sit on top of LLMs.

We are already seeing the core OpenClaw concepts propagate into the frontier companies' offerings.

Anthropic keeps shipping these
  • Claude DispatchLaunched March 17, 2026

    Text your desktop AI from your phone and come back to finished work. This is Anthropic's direct answer to OpenClaw-style delegation.

  • Claude CodeOngoing

    Adding skills, agents, and dispatch features that mirror OpenClaw's architecture.

  • Other frontier labsFollowing the playbook

    Same script — persistent agents, skill registries, multi-channel delegation.

The Career Angle

By starting to use OpenClaw yourself, you are going to get ahead of the curve of where these companies are going. For those working in bigger, established organizations or companies who are just now trying to figure out how to adopt AI — this is going to be a huge leg-up.

The Bottom Line

You will be thesmartest onein the room.

Chapter TAKEAWAYSClosing

Key takeaways.

01

Step function

OpenClaw is a step function in AI productivity — delegation, not just prompting.

02

A VM is fine

You don't need expensive hardware — a VM on your machine works.

03

Zero-trust

Security is non-negotiable — isolate everything.

04

Pattern match

The concepts transfer everywhere — pattern match across tools.

05

Start small

Start with the basics (one cron, one heartbeat, one skill), then expand.

06

Get ahead

The frontier labs are following OpenClaw's lead — learning it now puts you ahead.

Thanks for listening · questions now