Articles

An AI agent with nothing to install

Many developers, architects and consultants work on a machine where they may not install anything. An AI agent does not have to run there. It can run somewhere else and be reached from a browser tab, and then the question is where.

An agent is not a chat

A chat answers a question. An agent works: it reads files, writes code, runs commands, looks at the result and corrects itself until the task is done.

That is why it matters where the agent runs. A chat can at worst give a bad answer. An agent has the rights of the machine it runs on: the files, the keys and the network.

Most coding agents are programs you install on your own machine and run in a terminal. On a locked work machine it stops there.

Three places an agent can run

With a local agent. The agent is a program on your own machine, and you must be allowed to install it. It works with your user's rights, so what you can reach, it can reach. The tools can ask permission before each command; how much they may do without asking is a setting.

In a sandbox of its own. A machine that is only for the agent, at a provider you have chosen yourselves. The agent can reach nothing but what you put in there: nothing has to be installed, your own machine carries no load, and the agent cannot break anything on it. That is what we have built with Nomadics AI: a sandbox at Scaleway in Amsterdam with OpenCode, an open source coding agent, which opens in a browser tab.

In the vendor's cloud. Claude Code on the web clones your repository to a virtual machine that Anthropic runs. GitHub Copilot's coding agent works in an environment built on GitHub Actions. Nothing has to be installed, but it takes an account with the vendor and access to your code, and both vendors are American companies. At Anthropic the environment can be set up with a script, but it is built around GitHub, the machine is taken back when it stands idle, and the models are Anthropic's own.

The agent does not have to run on your machine

Three places a coding agent can run, and what sits beside it. The arrows show where what you write is sent.

With a local agent E.g. Claude Code or OpenCode in your terminal.
Your machine The agent Your filesYour keysYour network
The model in the vendor's cloud
  • Against: You must be allowed to install the agent and its tools
  • Against: Uses your machine's power
  • Against: The agent can reach your files, keys and network
  • For: Your own environment with everything you have installed
In a sandbox of its own E.g. Nomadics AI.
Your machine BrowserSSHFolder sync
The sandbox Scaleway, Amsterdam The agent Only what you put in
The model Scaleway, Paris
  • For: Ready at once: nothing to install on your machine
  • For: No load on your machine
  • For: Cannot break anything on your machine
  • For: A remote environment of your own, which you build out yourself
  • For: From the trial account: your own terminal and Claude Code
Try an agent
In the vendor's cloud E.g. Claude Code on the web.
Your machine Browser
The vendor's cloud Anthropic or GitHub, American companies The agent A copy of your codeThe model
  • For: Ready at once: nothing to install on your machine
  • For: No load on your machine
  • For: Cannot break anything on your machine
  • Against: The machine is taken back when it stands idle
  • Against: Needs GitHub and a paid subscription
  • Against: The vendor's own models only
Drawing: Nomadics. The points for the vendor's cloud apply to Claude Code on the web, from Anthropic's own documentation, read 18 November 2026. The sandbox is Nomadics AI.

Where what you write ends up

Three things leave your machine when you use an agent: what you write to it, the files it reads, and the answers the model gives. Ask the same three questions of any tool: where is the machine, where does the model run, and which country's law applies to the company behind it?

For Nomadics AI the answers are: the sandbox is at Scaleway in Amsterdam, the model runs at Scaleway in Paris, and Scaleway is a French company. The model is DeepSeek V4.1 Flash, an open model of Chinese origin. The origin is Chinese, the operation is French; we say both.

In the free try the sandbox has no access to the internet. The agent can write and run code, but it cannot fetch anything from outside and cannot send anything out.

What it is not for

A sandbox at someone else's is still at someone else's. Do not put customer data, personal data or trade secrets in a free try, nor in a paid tool, before your company has decided which data may be sent to which models.

That decision is far from made everywhere. In 2026 Stack Overflow asked some 13,000 developers what best describes the use of AI at their workplace. 24 % answered that the organisation provides approved AI tools and guidance. For 46 % it is mostly up to the individual or the single team.

One developer in four has approved AI tools and guidance

What best describes how AI tools and agents are used in the developer's organisation today. One answer each.

Approved tools and guidance
24 %
The organisation provides approved AI tools and guidance for how to use them 24 %
Up to the individual or the team
45.6 %
Mostly optional and left up to individuals 29.7 % · Some teams encourage it, practices vary 15.9 %
Expected or measured
21.7 %
Expected of some employees for certain work 11.5 % · Measured as part of performance 10.2 %
No AI, or not sure
8.7 %
The organisation does not use AI tools 6.1 % · Not sure 2.6 %
Source: Stack Overflow, 2026 Developer Survey, the question “AI at work”, 13,167 answers from developers worldwide. Not representative. The grouping is ours.

Where the decision has not been made, the answer is often that nothing may be installed. A try in an isolated sandbox is a way to see what an agent can do, with a task that holds nothing sensitive, while the decision is being made.

The agent's answers can be wrong. Read the code before you use it.

Try it in a quarter of an hour

Type a task, and the agent starts in a fresh sandbox. No account and nothing to install. The try is open from the Nordic countries and lasts 15 minutes, and the work is kept a week.

Three tasks that show the difference between an agent and a chat: ask it to write a script and run it. Paste a piece of code with a bug, and ask it to find the bug and prove the fix with a test. Ask it to build a small page and explain how the files fit together.

If the agents are to run in your own cloud account, with your policy for data and models and with an audit trail, that is a different job. We help with that too.