Imran was restarting OpenClaw's gateway once an hour.
Not once a week. Once an hour. On a bad day, he spent more time babysitting the setup than actually using it.
So he switched. Three weeks later he's running his whole life through Hermes Agent, his token bill dropped from $130 every 5 days to about $10, and he installed the thing on a cheap Android phone.
Everyone's asking the Hermes Agent vs OpenClaw question right now. Greg Isenberg brought Imran (he works at Alif, a venture fund) on the Startup Ideas podcast to answer it with receipts.
I pulled the full breakdown apart. Here's what actually matters for you as a founder.
The 3 problems that made him quit OpenClaw
Imran didn't leave OpenClaw because of hype. He left because of three specific failures:
→ Memory was not working reliably in his setup. That is different from OpenClaw lacking memory: its official documentation describes persistent Markdown memory. Imran reported having to repeat himself.
→ The gateway kept dying. One day he restarted it every hour. His words: "I spent more time setting up OpenClaw than actually using it to make my life better."
→ Token burn with zero visibility. Money was disappearing and he couldn't see how or why.
If those problems sound familiar, compare your own logs and spending before deciding which part of the setup needs to change.
He tried Nebula first (his take: better if you just want an AI co-worker). But for personalized workflows and a system that learns over time, he landed on Hermes Agent. He's been on it 3+ weeks now, which in this space is basically a long-term relationship.
What Hermes actually does differently
Hermes addressed the problems Imran described in his own setup. That is a useful case study, though it is not a controlled comparison of both products.
Hermes keeps curated long-term memory in MEMORY.md and USER.md, with conversation history stored separately in SQLite. Its memory documentation distinguishes those two systems. Imran described retrieving information from an earlier conversation; that does not mean every completed task is automatically added to long-term memory.
40+ tools out of the box. Browser, web search, cron jobs, image generation, even home assistant stuff. No hunting through a skills marketplace on day one. On Mac it ships with Apple Notes, Reminders, Find My, and iMessage skills pre-installed.
It stays up. Imran hasn't restarted it in over a week. After the once-an-hour gateway days, that alone justified the switch.
Install is one command on Mac, Linux, or WSL. If it's your first dev tool on a Mac, run xcode-select --install first and you're in.
Keep backups and review release notes before updating an agent that can act on your accounts. The episode described a fast-moving project and an older backup installation; it does not establish how long an update will take on your machine.
The 90% bill cut (this is the part that matters)
Here's the number that made me sit up.
Imran reported that token spending for his workflows fell from roughly $130 every five days to about $10 over the same period. He changed both model choices and how repeat tasks ran, so that saving cannot be attributed to the agent software alone.
Two moves did it:
Move 1: compare model prices through OpenRouter. The episode cited Qwen 3.6 Plus at about $0.33 per million input tokens and mentioned a temporary free model. Treat those as prices quoted in the recording. Check the current model, provider, input and output rates before estimating your own bill.
Move 2: turn suitable repeat tasks into scripts. A report or digest may not need an LLM on every run. Ask the agent to write and test the code, then check what the script calls. It can avoid recurring model tokens when it makes no model requests, but hosting, other APIs and maintenance can still cost money.
If you come from software, this is just "don't repeat yourself." If you don't, think of it as paying for the work once instead of every single morning.
People are spending hundreds of dollars a day on agent setups right now. Greg said it on the episode and nobody blinked. At founder scale, a 90% cut on that line is real payroll money.
He runs it on an Android phone (seriously)
Imran's favorite agent, Cookie Monster (all his agents are named after Muppets), doesn't run on a server. It runs on a Solana Seeker phone running Android 15, sitting on his desk.
The setup: an app called Termux (a terminal for Android), plus the Termux API app from F-Droid. That second one is the unlock. It gives the agent access to the phone's actual hardware: battery, Wi-Fi, camera, brightness, vibration, notifications, screen taps.
Why bother?
→ An Android phone you already own may be a practical test device. Compare its reliability, permissions and power needs with the hardware you would otherwise use.
→ Put a SIM card in it and your agent can read incoming SMS, which means it can handle two-factor logins on its own.
→ Imran described direct posting from the phone as a project he wanted to try. That plan is not evidence that a device MAC address improves reach or bypasses platform limits.
An always-on, low-power, dedicated agent box for the price of a used phone. That's a real edge while everyone else fights over Mac Minis.
What he actually automates day to day
The demos are fun. The daily wins are the point.
Email triage. Every morning an agent goes through his inbox, deletes the junk, unsubscribes from dead newsletters, and hands him a digest of what matters. Saves him 30 to 60 minutes a day.
Dinner. He sent one 8-minute Telegram voice note walking through every item in his fridge and pantry. Now he gets three recipes a day matched to his ingredients and fitness goals. Small thing. Removes a decision.
A self-building dashboard. After about 20 days of steady use, his Hermes agent built him an Obsidian setup on its own: a home file with this week's priorities, today's tasks, upcoming travel, work and personal split out. Refreshed automatically every morning. He never sat down and designed it.
His actual job. Imran works at a fund. With the background work automated, he talks to 20 to 30% more founders. Better signal, better deal flow. If you value your time at $500 an hour, do that math on your own calendar.
I've covered what founders are already building on the OpenClaw side in OpenClaw use cases to automate your business. The playbook transfers almost 1:1.
The prompts that make it click
The install takes 5 minutes. Knowing what to ask is the actual skill.
Imran's starter list, straight from the episode:
→ "What have I been procrastinating?" (it has your to-do list, it knows)
→ "What's the most important thing to work on today?"
→ "What tasks am I doing every day that I should automate?"
→ "What's one tool you could build me tonight that would make tomorrow easier?"
→ "Is there anything important today that I missed?"
Try the fourth prompt with a small task, then test the result before relying on it. A useful tool still needs a clear job and a way to check its output.
His self-reflection experiment drew on ELIZA, the 1960s program created by MIT computer scientist Joseph Weizenbaum. MIT describes its history. Turning a description into an agent skill may inspire a reflection prompt, but it does not reproduce the original program or establish a clinical benefit.
He also ported in G stack, Garry Tan's free skill that packages the Y Combinator weekly-iteration process. If you're building a startup outside SF, that used to be knowledge you had to get into YC to absorb. Now it's a file you bolt onto your agent.
So, should you switch?
Greg asked the direct version: what stops me from installing Hermes and crawling back to OpenClaw a week later?
Here's the honest breakdown.
Stay on OpenClaw if it's stable for you, your token bill doesn't hurt, and you've already built your workflows around it. Switching has a cost too.
Test Hermes if your current setup keeps failing or costs more than the work justifies. Run a few representative tasks through each setup, record model and hosting costs, and compare the time you spend fixing problems.
Either way, one rule from Imran that I'd frame on the wall: "Customizing it is not the skill." He compared this whole scene to 90s tuner car culture. Some people polish the build forever and never drive anywhere. The agent only matters if it buys you back hours and moves your business.
Pick one, spend 7 days actually routing your work through it, and judge on saved hours.
If you want the deeper setup walkthrough on the Hermes side, I broke that down in Hermes Agent desktop: full setup and real use cases.
FAQ
Is Hermes Agent better than OpenClaw?
Imran preferred Hermes for his setup. Both products have memory features, and his experience alone cannot establish which is more stable for most founders. Compare the workflows, model costs and reliability that matter to you.
How much does Hermes Agent cost to run?
The software is free, but model requests and hosting can cost money. Imran reported spending about $10 per five days after changing models and turning repeat tasks into code. Your total depends on the tasks, selected providers and infrastructure; check current rates before budgeting.
Can you run Hermes Agent on an Android phone?
Yes. Install Termux (a terminal app), add the Termux API app from F-Droid, and run the Android install script. The agent then gets access to the phone's camera, battery, notifications, and SMS. With a SIM card in, it can even handle two-factor codes. It's a cheap always-on agent device.
Is Hermes Agent safe to install?
Review the permissions an agent needs before giving it access to accounts, files or devices. Keep backups, protect credentials and review updates. An agent auditing its own setup is not an independent security assessment. A container can help separate files and processes, but it does not make every action inside it safe.
One episode like this saves you the 3 weeks of trial and error Imran already paid for. That's the whole idea behind the Profitable Founder Podcast: bootstrapped founders opening up their real numbers and the stacks behind them.
