Most people who want a real AI agent end up in the same trap. They spin up something powerful, then spend their weekends babysitting it instead of using it. The setup eats your time. The maintenance eats the rest.
I just watched a walkthrough that flips that whole script. The creator behind it, Matthew Berman, spent months running OpenClaw on a dedicated machine, and he’s refreshingly honest about the cost. It was expensive. It was fiddly. He admits he spent more time keeping it alive than actually getting value out of it. Then he tried Perplexity Computer, and his takeaway is simple: for most people, this hosted version is the better call.
Let me break down what he showed, because the contrast is the whole story.
The old way vs the new way
Here’s the core idea. Perplexity Computer is basically OpenClaw, but fully hosted. Same powers, none of the server babysitting.
Think about what the self-hosted route actually asks of you:
- Install it on a separate machine for safety
- Hand over your own credentials to each service
- Manually wire up every connection, one API key at a time
- Keep patching, updating, and fixing it when something breaks
The original poster did all of that. He’s technical, he likes getting his hands dirty, and even he calls it a grind. The hosted version skips every one of those steps. You get a full AI agent that writes code, runs code, searches the web, and uses frontier models, without owning the plumbing.
The trade-off he’s clear about: if you want total control and everything running locally, self-hosting still wins. For everyone else, hosted is the smarter default.
What stood out
A few things made the expert lean toward Perplexity Computer.
Threads by default. With the self-hosted setup, he had to manually rig up multiple topics so the agent wasn’t mixing context between unrelated tasks. Perplexity is threaded out of the box. Each task is its own thread with clean context, and they all run in parallel. He could fire off one task, click away, and start another while the first kept working. Threads can even delegate to sub-agents underneath.
Connectors. This is the part he spent the most tokens and frustration on before. Gmail, Calendar, Telegram, Drive, Notion, GitHub, Dropbox, Linear, Box, OneDrive, and on and on. Hundreds of them, pre-built. You click, you authenticate like a normal login, and you’re done. No juggling API keys for everything. And because it’s wrapped in Perplexity’s security, that old “is this even safe?” worry about handing credentials around mostly goes away.
Skills. Just like Claude Code or the self-hosted agent, you can teach it how to do things. It even ships with a “create skill” skill, so you describe what you want and it builds it.
Model choice. You pick the orchestrator model: Opus, GPT, or Sonnet. The creator mentioned he runs a split between Opus and Sonnet, and called Sonnet a beast that costs far less.
The use cases he actually ran
This is where it got real. Every one of these is something the author had already built the hard way, then rebuilt here in minutes.
- UFC fight brief. He asked it to gather the fight lineup, odds movement, weigh-in results, interviews, and injury chatter, then deliver a compact brief Saturday afternoon. It set the whole thing up, then paused itself and scheduled a wake-up for 1pm Saturday to do a final research sweep.
- Food journal. He recreated his stomach-troubleshooting food log in seconds. Snap a photo of a burger, and it logs the time, ingredients, and details into a markdown file.
- Model benchmark chart. He had it pull fresh benchmark data on a new model release and build a comparison chart against rival models. It caught its own mistake mid-task, re-searched, ran sub-agents in parallel, and produced a clean visual.
- Earnings preview. Every Sunday it scans the week’s quarterly earnings, knows his interests lean tech, asks which ones he wants, then pulls and analyzes transcripts right after each call.
- Knowledge base. The big one. He asked it to build a persistent web app that ingests links, articles, and videos, converts them to embeddings, and lets him search in natural language. It planned with one model, delegated the build to another, and shipped a working full-stack app in about five minutes.
Proactive results land in Perplexity’s mobile app, which he praised, or you can route them to Telegram. Adding Telegram is one click, no hunting for bots.
How to try it yourself
If you want to follow the path the creator laid out:
- Get a Perplexity plan (this also covers search and the rest of the suite).
- Open Perplexity Computer and start a task, which acts as a clean thread.
- Flip on the connectors you need, like Gmail and Calendar, with a quick login.
- Pick your orchestrator model based on cost and muscle.
- Use the “create skill” feature to bottle up repeat workflows.
One honest note on cost. It isn’t free, and neither was the self-hosted route. You pay monthly for the plan, get a pile of credits for Computer, then pay for usage on top. The dashboard shows credits per task, so you can see what’s cheap and what’s pricey before it surprises you.
The creator’s verdict was the part that stuck with me. He rebuilt his most valuable, most complex workflows without touching security headaches or burning tokens on setup. That’s the real unlock here.
Want the full screen-by-screen tour and his exact prompts? Watch the original video, it’s worth your time.