Inside the OpenClaw Workshop: How AI Agents Are Quietly Running Parts of Emarketed

Today I hosted an OpenClaw workshop, a deep dive into how we’re using open-source AI agents at Emarketed to handle lead generation, content publishing, website updates, and a lot of the small, repetitive work that used to live on a human to-do list. If you couldn’t make it, here’s the full recap.

View the Workshop Slides

OpenClaw Workshop Slides

View the full May 26, 2026 workshop slides

What Is OpenClaw?

OpenClaw is an open-source AI agent platform. The short version: it lets you spin up persistent, model-agnostic agents that live inside the messaging app you already use, Slack, Telegram, WhatsApp, iMessage, or in my case, Discord.

A few things make it different from just chatting with ChatGPT or Claude:

It’s model-agnostic. You can run Claude, GPT-5, Gemini, DeepSeek, or pretty much anything else. Different agents on your team can run different models for different tasks.
It has persistent memory. Each agent reads its “soul file” (who it is), a user file (who you are), and yesterday’s and today’s memory at the start of every session.
It has a heartbeat. By default, every 30 minutes the agent wakes up, checks for scheduled jobs, and gets to work. You can tune that cadence, five minutes, ten minutes, but the faster it ticks, the more tokens you burn.
It’s only about five months old. Which means it’s improving fast, but it still acts like a smart intern. More on that below.

How My Agent Team Is Structured

I run several agents in Discord, each with a specific role:

Main, my general-purpose agent, currently running on GPT-5.5
Forge, coding and website updates
Nova, writing and content
Recon, sales ops and prospecting

Some people prefer to run a single agent that spawns sub-agents on the fly. I prefer having dedicated agents I can talk to directly. If Forge is in the middle of a website update, I can still hop over to Nova and kick off a blog draft without waiting.

Each agent gets its own channel for reports, agent updates, sales leads, AI news summaries, which keeps Discord from turning into chaos.

Where OpenClaw Actually Lives

You’ve got two options: run it locally or run it in the cloud.

Running it locally on a dedicated Mac Mini is fine if you’ve got a spare machine and aren’t worried about exposing your main computer. I run mine in the cloud on exe.dev. About $20 a month gets you twenty virtual servers, and every instance comes with its own AI assistant named Shelly. When I need to update OpenClaw, free up disk space, or run a security audit, I just tell Shelly. I’ve never once had to talk to a human at exe.dev to get unstuck. If you want to try it, message me and I can get you a free month.

One rule I learned the hard way: don’t run OpenClaw on an unrestricted API key. I once spent $400 in a couple of hours when an agent got stuck in a loop. Now I either run it on a monthly subscription, currently the $100 a month ChatGPT plan since Anthropic discontinued Claude subscription access for agent platforms, or use OpenRouter with hard spending limits.

What We Use OpenClaw For at Emarketed

Here’s where the rubber meets the road. A few things my agents handle every day without me touching them:

Lead generation. Every day, Recon goes out and finds companies in our target niches. It pulls domains, runs them through Hunter.io to find and verify emails, and drops the validated contacts into SalesHandy, which kicks off a three-email cold sequence. Hands-off prospecting.

Content publishing. One of our agents researches trending AI and marketing news every day, writes a blog post in our voice, generates the images, links to sources, and internal-links to target keywords on our site. This has been running for nearly three months. Our traffic and rankings for terms like AI visibility and answer engine optimization agency have both climbed since we turned it on. It’s essentially a full-time SEO writer that doesn’t sleep.

Website updates. Our sites live on GitHub and deploy to Netlify. The agent has access to the repos, so when an event wraps up and I want the page updated, I just message it on Discord: “swap the date, add the recap, drop in this photo,” and it pushes the change. Gone are the days of pinging a developer for a five-minute edit.

Social posting. We use Metricool for social scheduling, and because Metricool has an API, the agent can drop scheduled posts directly into the calendar across Facebook, Instagram, and LinkedIn. Anything with an API is fair game.

Client content. Each client gets its own knowledge base, writing style, brand voice, about-the-company doc, so content stays unique and on-brand for each account.

OpenRouter: The Underrated Power Move

If you’re going the API key route instead of a subscription, OpenRouter is the play. Pay-as-you-go, hard spending limits, and access to roughly 400 models. You can route different agents to different models, DeepSeek for cheap drafting, Claude for nuance, GPT for the main agent, and keep costs manageable.

What I just discovered today: OpenRouter now lets you spawn OpenClaw and Hermes agents directly within their platform. You pick your cloud provider, AWS, Digital Ocean, Hetzner Cloud, and spin it up under one umbrella. I’ll be doing a follow-up walkthrough on this soon, particularly to see if I can run a Descript agent in the cloud instead of locally on my laptop for transcription.

OpenClaw vs. Hermes

I’ve been running Hermes alongside OpenClaw for about a month. They’re functionally similar. Both let you run agents on the messaging app of your choice, both are model-agnostic, both handle scheduled tasks.

The main differences I’ve noticed:

OpenClaw has the bigger, more supportive community. Hermes folks are louder on X, but OpenClaw has more depth.
Hermes hallucinates a bit more. I’ve gotten the occasional gibberish Discord message that OpenClaw never produces.
Hermes leans on scripts. By default it tends to build a script to execute tasks rather than following step-by-step instructions every run. That makes processes more reliable, less chance of a step getting fumbled.

A fun example of what Hermes does for me personally: every morning and afternoon, it checks the MLB API and tells me whether there’s a Dodgers home game so I can dodge the traffic. It also monitors a 24/7 AI-generated news radio station called Hype Radio, summarizes what’s trending, and at the end of the day gives me a take on which stories matter for Emarketed.

Memory Is the Weak Spot, But It’s Improving

The one thing that isn’t perfect out of the box is memory. OpenClaw reads its soul file, user file, and recent memory on every heartbeat, but if you want it tight, you’ll want to look into community projects like Gbrain, a more optimized memory layer that I recently installed. The OpenClaw community is actively iterating on this, so expect rapid improvement.

A Few Hard-Earned Lessons

If you’re going to start running agents, a few things I’d tell my past self:

Treat it like an intern. It’ll make mistakes. It’ll time out. It’ll occasionally need babysitting. But it learns, and the wins compound.
Don’t run it on your main machine. Cloud or a dedicated box.
Cap your spending. Subscription plan or OpenRouter with limits. Never a raw, uncapped API key.
Keep an eye on API keys and credits. A blog pipeline can quietly break because an image-generation API ran out of credit. You can build another cron job that monitors those, agents watching agents.
Offload the right work. I’m not using OpenClaw to build full dashboards, that’s still Claude Code on my laptop. I’m using it for the repetitive stuff: email triage, daily research, scheduled posts, site updates, prospecting. Things I’d otherwise have to context-switch into.

Where I’m Headed Next

The piece I want to solve next is being able to work even when I’m not at my laptop. Hosting an event at Ground Floor and need to update the site? Message Discord from my phone, done. Driving somewhere and want to kick off a research task? Voice call to my agent. That’s the workflow I’m building toward, fewer dead minutes, more leverage.

If you want to see this stuff in action or compare notes, come to the next OpenClaw LA meetup at Ground Floor. And if you’d like help wiring up agent-powered workflows in your own business, whether that’s content, SEO, paid ads, or prospecting, that’s exactly what we do at Emarketed through our managed agent service. Get in touch.

Matt Ramage is the founder of Emarketed, a Los Angeles digital marketing agency, and the organizer of OpenClaw LA.