A repetitive task across several tools
Checking emails, updating a spreadsheet, preparing a report every morning: an agent can chain these steps without anyone stepping in.
OpenClaw · Hermes Agent · NVIDIA NemoClaw
Autonomous agents: a secure deployment, with clear limits.
An autonomous agent does more than answer: it reads files, browses the web, sends messages and runs commands. We install OpenClaw, Hermes Agent or NVIDIA NemoClaw in an isolated environment, with limited access and human approval for every important action.
01 Understand02 Decide03 Build04 Verify05 Measure
NVIDIA OpenShell
OpenClaw
OpenClaw
01Network
02Files
03Credentials
04Approvals
When is it useful?
An agent installed without rules
An employee is trying OpenClaw or Hermes with their own access. You want to take back control before an incident.
An agent with no data leaving Switzerland
These tools are open source: they can run on your servers, with a model hosted in Switzerland.
The tools we deploy
Three open-source tools, chosen according to the task. We do not modify them: we install and control them.
OpenClaw
An agent that takes requests through messaging apps and acts on the computer it is installed on: files, browser, commands.
OpenClaw security ↗Hermes Agent
An agent from Nous Research that learns: it writes its own procedures and keeps a memory from one session to the next.
Hermes Agent repository ↗NVIDIA NemoClaw
An NVIDIA framework that runs OpenClaw in a closed environment, with rules on network, files and access keys.
NVIDIA NemoClaw documentation ↗The risks, plainly
It can act
A useful agent runs commands and calls other systems. Every permission needs a limit and an owner.
It obeys what it reads
An email, a web page or a document can contain hidden instructions. The agent must not be able to do everything on that basis alone.
It holds keys
A password or access key left inside the agent turns a simple mistake into an incident.
It learns
Hermes writes its own procedures. A learned procedure is reviewed before it is reused.
What we deliver
Access inventory
Who can give the agent instructions, what it can read, change and send, and what needs human approval.
Isolated environment
A dedicated server or machine, a closed environment, network access limited to the destinations needed, separate files.
Dedicated accounts and keys
Accounts for the agent alone, with minimal rights, rotated keys and no personal accounts.
Testing and monitoring
Misuse tests before go-live, an action log, a shutdown procedure and a quarterly access review.
For an SME: how we work
An autonomous agent is an Engineering project: scope, price and timing are agreed before work begins.
AI Engineering
Secure installation, integration with your tools, testing and go-live. Follow-up can then run through a support subscription.
Fixed price on proposal, or subscription from CHF 1,800/month (10 h) excl. VAT
See the offer AI Engineering →Frequently asked questions
Do you really need an autonomous agent?
Often not. A conventional automation or an assistant such as Microsoft 365 Copilot is enough for many tasks. We check this before proposing an agent.
Is NemoClaw enough to secure OpenClaw?
No. According to NVIDIA’s documentation, NemoClaw protects the system: closed environment, network, files. The agent’s own rules, such as who can talk to it and which actions need approval, still have to be defined. We handle both.
Does the data stay in Switzerland?
The agents run wherever you install them. What matters next is the model used and the services the agent calls: we choose them explicitly.
Can several people share one agent?
OpenClaw is designed for a single trusted user. For a team, we separate agents or environments rather than share one access.
Is Numezis a partner of OpenClaw, Nous Research or NVIDIA?
No. Numezis does not claim any partnership or certification. These are third-party technologies that we deploy independently.
A first conversation to agree the next step.
Describe your need in a few lines. We reply within two working days with the offer that fits.