Hermes, Agent Zero or OpenClaw: which self-hosted agent should you run?

"Self-hosted agent" covers three quite different products. One is a service that answers your team's messages, one is a whole desktop for an agent to work in, and one is a coding agent that lives in your terminal. Choosing between Hermes Agent, Agent Zero and OpenClaw starts with what you want it to do.

At a glance

	Hermes Agent	Agent Zero	OpenClaw
Built for	Always-on business work	Open-ended tasks on a full computer	Coding and personal automation
Where it runs	A service on a host you control	A Docker container with a Linux desktop	Your machine, in the terminal
How you talk to it	Messaging apps, desktop app, CLI	Web interface and desktop	Terminal
Isolation	Your own host; approvals for risky actions	The container boundary	Permissioned tools; sandboxed skills
Models	Provider-agnostic	Configurable	Model-agnostic
Memory	Persistent; learns skills	Per-project memory	Persistent across sessions
Main risk	Reading untrusted inboxes	Host folders you mount	Broad local access

Choose Hermes Agent when the work is a queue

If the job is "watch this mailbox, form or channel and act on what arrives", Hermes is the natural fit. It runs whether or not anyone's laptop is open, speaks to many messaging platforms and can run scheduled jobs. It is aimed at businesses with a repetitive intake problem rather than at developers. The first fortnight needs real supervision, for the reasons we set out in the small-business case.

Choose Agent Zero when the work is open-ended

If the task is research that ends in a document, a browser-heavy chore or a project where you want the agent to build its own tooling, the whole-desktop approach pays off. The container protects your machine, and per-project isolation keeps clients apart. It is the least predictable of the three and the wrong choice for a production service.

Choose OpenClaw when the work is code

If you want Claude Code-style ergonomics without the lock-in, and you are comfortable owning the isolation decision, OpenClaw is the closest to a conventional coding agent. It suits polyglot and cost-sensitive work, and a large skill library comes with a supply-chain caveat: read what you install.

Rule of thumb

Queue of small jobs: Hermes. A messy one-off that needs a computer: Agent Zero. Software in a repository: OpenClaw. If two of those describe you, run them separately rather than making one tool do both.

What they share

All three are community or open-source projects with no support contract, all three need you to own uptime, updates and key rotation, and all three move fast enough that any comparison, including this one, ages. Whichever you pick, give it its own scoped accounts, keep irreversible actions behind approval, and log what it does somewhere a person reads.

For a fuller feature-by-feature view, the spec sheets on the harness board and framework board are kept current, and Anatomy of an agent harness explains the pieces being compared.