Hermes Agent: The New OpenClaw?
Imran walks through how to install and use Hermes Agent, positioning it as a superior alternative to Open Claw due to its built-in memory system, stability, and token cost efficiency. The episode covers installation on Mac and Android, connecting with tools like Obsidian and G Stack, creating skills, and practical use cases for personal and professional automation.
Summary
The episode features Imran explaining Hermes Agent to host Greg, framing it as a potential replacement for Open Claw. Imran outlines three core problems he experienced with Open Claw: no persistent memory requiring constant re-instruction, frequent gateway restarts (sometimes hourly), and opaque token spending. Hermes Agent addresses all three by providing a built-in memory system that writes successful task completions to a SQLite database, improved stability (no restarts needed in over a week), and clear model/cost visibility via Open Router integration.
Imran demonstrates the installation process, which is a single command on Mac, Linux, or Windows Subsystem for Linux, and highlights that Hermes ships with 40+ built-in tools covering browsing, web search, cron jobs, image generation, and home assistant capabilities, along with pre-installed Apple ecosystem skills (Notes, Reminders, iMessage, Find My) on Mac.
A significant portion covers Hermes on Android via the Termux app, which gives the agent access to phone sensors, camera, Wi-Fi, SMS, and device controls through the Termux API. Imran suggests this enables cheap, always-on agent hardware as an alternative to sold-out Mac Minis, with potential applications in social media automation that posts natively from a device to avoid API reach penalties.
On cost optimization, Imran reports reducing his token spend from roughly $130 every five days to about $10 every five days — a 90%+ reduction — by switching to Hermes with Open Router and converting recurring tasks into deterministic code rather than re-running LLM inference each time.
The conversation covers agent design philosophy: Imran runs four agents named after Muppets, but recommends most people maintain one or two (personal and work). He uses Obsidian as a clean frontend for agent-generated markdown dashboards, replacing walls of Telegram text. He also discusses G Stack by Garry Tan (Y Combinator's startup methodology encoded as an agent skill) and a self-built psychiatrist-style skill based on Joseph Weizenbaum's ELIZA program. Imran closes by emphasizing that customizing the agent is not the core skill — using it to actually get work done is.
Key Insights
- Imran argues that Open Claw's lack of persistent memory meant users had to re-explain the same instructions repeatedly, and its gateway required restarting as often as once an hour, making setup overhead outweigh actual productivity gains.
- Imran claims Hermes Agent uses a SQLite database to log all past tasks, enabling it to retroactively search conversation history for things like API keys that were mentioned but never formally saved to environment variables.
- Imran reports achieving a greater than 90% reduction in token spend — from approximately $130 every five days down to $10 every five days — by switching to Hermes with Open Router and converting recurring tasks into static code instead of repeated LLM inference.
- Imran demonstrates that Hermes Agent can be installed on an Android phone via Termux, and that adding the Termux API app grants the agent access to phone sensors, camera, SMS, Wi-Fi controls, and the vibration motor, effectively creating a low-cost always-on agent device.
- Imran suggests that posting social media content directly from an Android device running Hermes — rather than through a scheduling API — could preserve algorithmic reach because the post appears to originate from a real device with a genuine MAC address.
- Imran describes using a local speech-to-text model to send an 8-minute Telegram voice message cataloguing his fridge and pantry contents, then having the agent generate daily recipe suggestions based on available ingredients and fitness goals, which he says removes significant mental load.
- Imran explains that G Stack — built by Y Combinator CEO Garry Tan originally for Claude Code — encodes YC's week-over-week startup improvement methodology as a free, installable agent skill, making previously exclusive accelerator-style guidance broadly accessible.
- Imran argues that learning to use Hermes Agent is not the actual skill — customizing and tinkering with the agent is a distraction, and the real value is in what work gets completed with it, citing his ability to engage with 20–30% more founders at his fund as the tangible outcome.
Topics
Full transcript available for MurmurCast members
Sign Up to Access