Runtime guide

Hermes vs OpenClaw: choose the runtime after you understand the job.

Hermes and OpenClaw can both run self-owned agents, but they organize agent state and operation differently. The best choice depends on the destination you already operate, the channels and tools you need, and the isolation you expect—not on a generic winner.

Short answer

Choose Hermes when its profile, gateway, skill, scheduling, and model setup matches the environment you want to operate. Choose OpenClaw when its workspace, gateway, channel, skill, memory, and heartbeat model fits your existing setup. If either is already running, preserve it and let the coding assistant inspect the host before recommending a separate profile, agent, or container.

How the operating model differs

Hermes

Official Hermes documentation describes profiles as separate Hermes home directories with their own configuration, environment file, memory, sessions, skills, cron jobs, and state database. A profile is useful state separation, but its tools can still share the host unless you add a stronger boundary.

Official Hermes profiles documentation ↗

OpenClaw

Official OpenClaw documentation describes the workspace as the agent's working directory and context home. It explicitly warns that the workspace is not a hard sandbox: absolute paths can reach elsewhere unless sandboxing is enabled.

Official OpenClaw workspace documentation ↗

What stays the same

A role that tries to own too many unrelated outcomes becomes unreliable in either runtime.
Memory needs a purpose, source, retention rule, and way to correct stale information.
Heartbeats or schedules need a bounded responsibility and a visible definition of done; frequent background runs also increase model usage.
A messaging channel is an access surface, not the definition of the agent.
App credentials, OAuth, model authorization, and destructive operations need to be configured in the customer-owned environment.

Five questions to choose safely

Is Hermes or OpenClaw already installed and carrying important state?
Do you want a separate state profile/agent or a stronger container boundary?
Which chat channel and business apps does this role actually require?
How will the model be authorized and paid for?
What must be tested before schedules or live actions are enabled?

Agent Factory's role

Agent Factory keeps the business plan portable, then formats the build handoff for the selected runtime. Codex or Claude Code performs the destination-specific inspection and implementation. That prevents the planning product from pretending it controls your host, tokens, channel access, or final deployment.

Not sure yet?

Start with the role and choose the runtime during review. The business outcome should survive either implementation path.

Plan a self-owned AI agent →