Guide
OpenClaw and Hermes Agent in business
Banning them doesn't work. Ignoring them doesn't either. Here is how to adopt them without opening a breach.
What we're talking about
OpenClaw and Hermes Agent are two open-source personal AI agents, both MIT-licensed. The first wires into messaging apps and runs tasks on your machine or server; the second is a continuous-memory agent runtime that learns from how you use it. Both rank among the most starred projects on GitHub, and we track their metrics on our Radar rather than freezing them into a page.
We are partners of neither, and we promise no features on their behalf: these projects ship several releases a month, and a feature list written today would be wrong in six weeks. This guide covers what doesn't change — how to bring them into a company without creating a breach.
It is written for the people who have to decide: IT and security leadership, DPOs, executives. Not for anyone looking for an install tutorial.
What these agents really change
A conversational assistant answers. An agent acts: it reads your messages, opens your documents, sends replies, triggers actions in other tools. That is a difference in kind, not in degree — and it is what moves the risk.
Three properties stack up here that rarely stack up elsewhere. The agent reads content written by third parties, so it is exposed to hidden instructions. It acts on your behalf, so its mistakes carry your signature. It remembers, so it accumulates personal data nobody decided to collect.
Each is manageable on its own. Together they call for a method — which is what our hardening framework sets out.
Three questions to settle before installing
Where does it run? On an employee's laptop, the agent inherits their session, their access and their secrets, and becomes a pivot point into the network. On a dedicated server you control, the scope is decided rather than inherited.
What can it reach? The default answer is "everything its user can see" — which is to say, too much. A dedicated service account with folder-by-folder permissions turns a question of trust into a question of configuration.
Who approves what? Reading is reversible; sending, paying, publishing and deleting are not. The line between what the agent does alone and what it proposes to a human is the most structural decision of the deployment — and the easiest one to postpone.
OpenClaw or Hermes Agent?
The honest answer: it depends on the use case, and the choice of tool is not what will sink the project. Both are MIT-licensed, so commercially usable with no strings attached; both move fast.
Where the choice genuinely matters is memory. Hermes Agent's continuous memory is the point of the product — and the part most exposed under GDPR: a memory you cannot search is a memory you cannot purge, and the right to erasure applies here as anywhere.
Our position on each, licence and governance included, is published on the Radar and updated automatically — not here, where it would go stale.
What if your teams already installed one
That is the most common case, and bans backfire without exception: they move usage out of your sight instead of removing it. It is the same mechanism as AI used behind your back, with more permissions attached.
The sequence that works: survey what is already running and with what access, replace personal installs with a governed deployment, write the usage policy, then train. In that order — a policy published before the survey describes nothing real.