bootkey.ai

research // measurements

How small is a small agent?

· 4 min read

bootkey now hosts three agents: Hermes Agent, PicoClaw and ZeroClaw. Before we offered the two small ones, we ran each of them on the machine every bootkey agent gets, a DigitalOcean box with 1 vCPU, 1 GB of memory and 2 GB of swap, and wrote down what they cost in memory and in time. We ran OpenClaw on the same box too, because “OpenClaw alternative” is how a lot of people start looking.

The numbers

OpenClaw 2026.11.3 Hermes Agent 0.23.1 ZeroClaw 0.9.0 PicoClaw 0.3.1
Written in TypeScript (Node) Python Rust Go
Licence MIT MIT Apache-2.0 MIT
Memory when idle about 430 MB, plus 750 MB pushed out to swap about 360 MB about 55 MB about 35 MB
One chat turn on the 1 GB box 206 seconds not measured here 3 to 4 seconds about 5 seconds

Every chat turn used the same model (DeepSeek V4.1 Flash through OpenRouter) and the same question: who are you, and what is this machine’s hostname? Check it with a tool. Each agent ran the hostname command and answered correctly. OpenClaw did too. It just took three and a half minutes, because the box spent most of that time moving memory in and out of swap. At its peak the machine had all of its memory in use and 1.2 GB in swap.

To be fair to OpenClaw: we sent its turn through its own command-line client, which starts a second Node process of about 400 MB. Talking to its gateway directly would save some of that. But the gateway alone, sitting idle, already needs more memory than the box has. OpenClaw wants a server with at least 2 GB, and 4 GB would be comfortable. We don’t host it today.

Tokens

ZeroClaw reports its own token use. Each turn sent about 17,800 input tokens before our question: its instructions and the descriptions of its tools. With a cheap model that costs very little, and our whole test run of PicoClaw and ZeroClaw (a dozen chats and a few scheduled jobs) came to under two cents. With a frontier model it adds up, so a fast, cheap model suits ZeroClaw. We haven’t measured the others this way yet.

What it took to host them

PicoClaw is one Go program with a JSON config and a separate YAML file for secrets. One thing caught us: an empty section written as channel_list: reads as null in YAML, and PicoClaw then refuses to start. Written as channel_list: {}, it starts. Its scheduler keeps UTC time, so duties on PicoClaw are set in UTC.

ZeroClaw is one Rust program that runs its gateway, heartbeat, scheduler and chat channels together. Its setup wizard needs a terminal, so we configure it with zeroclaw config patch, which takes a JSON Patch on standard input and stores keys encrypted. Once the program is running it owns its configuration, and changes go through its own local API instead. Its scheduler understands time zones, one-off jobs and running a job on demand.

OpenClaw has a good scripted setup: openclaw onboard --non-interactive with the key passed as an environment variable worked first time. Installing it from npm took 94 seconds and 651 MB of disk, and the gateway took 68 seconds to start.

Which one to pick

If you want everything bootkey can do, pick Hermes: job templates, coding hands, a browser, an email address, every chat app, and an agent that can decide for itself when to wake. If you want something small for a Telegram assistant or a few scheduled jobs, pick PicoClaw or ZeroClaw. ZeroClaw if you want duties in your own time zone. The agents page has them side by side.

How we measured

November 13, 2026, on DigitalOcean’s 1 vCPU, 1 GB machines with 2 GB of swap, Ubuntu, in New York and Frankfurt. Idle memory is the resident size of each agent’s own processes a few minutes after it started, read with ps; swap from free. Chat turns for PicoClaw and ZeroClaw went through the bootkey console and are timed on the box, from message to reply. OpenClaw ran under a separate user on a box whose own agent we had stopped, with Node 26 and the official npm package. Hermes was measured on a running agent with its coding hands installed.