Skip to main content
The Argon logo mark and wordmark inside the boxd card. Argon gives every agent run its own real MongoDB database, forked from your data. What the agent writes becomes a diff: you review it, merge what you want and discard the rest. Argon’s repository has an example that pairs it with boxd (pull request). The agent’s code runs on an isolated boxd machine and writes to an Argon sandbox. You get two sandboxes: boxd for the code, Argon for the data.

The example's README

The full guide: setup, credentials, the port check and clean up. Start there.

How it works

  • Two machines. MongoDB, the Argon console and host.py run on a data machine. The agent’s code runs on a second machine that host.py creates for each run with isolated and the argon-example network. It reaches the data machine at <data-machine>.boxd:27017 and no other machine you own.
  • No internet for the agent’s code. The agent machine installs PyMongo first. Then host.py sets its egress allowlist to none, so the code can reach MongoDB and nothing else. SSH is off.
  • A database user per run. The agent never gets Argon’s connection string. host.py creates a MongoDB user with readWrite on the sandbox’s database only, passes it on the command’s stdin, and drops it when the run ends.
  • Review first. A default run prints the diff and discards the sandbox, so your main branch stays as it was. --apply merges the diff of its own run.

Run it

1

Create the data machine and an API key

On the data machine, start MongoDB as a one-node replica set with authentication, install Argon (2.3.1 or newer, npm install -g argonctl) and start argon console. The README has the exact commands.
2

Check port 27017

The README’s check pings MongoDB from an isolated machine on argon-example and from one outside it. Only the first may answer:
3

Run the example

On the data machine:
The run took about 17 seconds, the agent machine included. host.py deletes the agent machine at the end.
4

Merge after review

Each run starts a new project, agent machine and sandbox, so --apply merges the diff of its own run.
agent.py stands in for code an agent would write, so no LLM account is needed. Replace it with your agent’s code and keep the rest.

Clean up

The data machine and the API key stay until you remove them:

Read more