OpenClaw2026.11.3
TypeScript (Node)
idle memory
≈ 1.2 GB
430 MB in memory, 750 MB pushed to swap
one chat turn
206 s
mostly swapping
needs 2 GB or more
openclaw alternative // measured, not guessed
OpenClaw made the idea popular: an agent with a soul file, a memory and a heartbeat, living on a machine you control. It also needs a big machine. We ran it next to the three agents we host, on the same 1 GB server, with the same model and the same question.
1 vCPU, 1 GB, 2 GB swap. DeepSeek V4.1 Flash via OpenRouter. November 13, 2026. Method and details.
OpenClaw2026.11.3
TypeScript (Node)
idle memory
≈ 1.2 GB
430 MB in memory, 750 MB pushed to swap
one chat turn
206 s
mostly swapping
needs 2 GB or more
Hermes Agent0.23.1
Python
idle memory
≈ 360 MB
with coding hands installed
one chat turn
–
not measured here
ZeroClaw0.9.0
Rust
idle memory
≈ 55 MB
one program, its own memory database
one chat turn
3 to 4 s
same model, same question
PicoClaw0.3.1
Go
idle memory
≈ 35 MB
one program, plain files
one chat turn
≈ 5 s
same model, same question
Every agent ran the hostname command and answered correctly, OpenClaw included. OpenClaw's turn went through its own command-line client, a second Node process of about 400 MB; its gateway alone already needs more memory than this machine has.
what you get // with any of the three
machine
Its own IP address, root over SSH, and the city you pick: Frankfurt, New York, San Francisco or Sydney.
mind
OpenRouter or Venice, on your own account and at their prices. Switch models whenever you like.
console
Chat, duties, its brain files with every version kept, logs, and a terminal in the browser.
care
Self-repair with a message to you if it stops answering, and hibernation for $0.50 a month.
questions
Not today. On the 1 GB machine every bootkey agent gets, OpenClaw needed 206 seconds to answer one question while the machine swapped. It wants at least 2 GB, and 4 GB would be comfortable. If you need OpenClaw itself, run it on a server that size.
Hermes Agent does the most of the three: chat apps, scheduled jobs, skills, a browser, coding, email and its own memory. PicoClaw and ZeroClaw share OpenClaw's idea of an agent that lives in plain files (a SOUL.md and a USER.md among them) in a much smaller program.
$9 a month for the agent's own machine, billed by the hour, with any of them. Model tokens go to your own OpenRouter or Venice account at their prices, with no markup.
bootkey doesn't import it for you. Because the agent's brain is plain files and you have root, you can copy your SOUL.md, notes and skills onto the box and edit them in the console.
Optional analytics
May we use Google Analytics to see which public pages people read? Nothing from your account, your console or your agents. Details