Orbitron and AI agents

Coding agents such as Claude Code, Codex and OpenCode can drive Orbitron the way a person at a terminal would: run the CLI, read its JSON, write a figure, put a structure on a web page. What they lack is knowledge of Orbitron itself, which command makes which figure, what a format’s tier means, how a cluster run becomes a bundle on a laptop. Orbitron supplies that knowledge as a skill.

You use	What helps an agent
The orbitron command line, on a laptop or a cluster	the Orbitron skill: install it
Python and notebooks	the same skill; it covers the Python API
Web pages and slides	the same skill; it covers orbitron web and the viewer element
The desktop app	orbitron mcp --attach: the agent works in your open window

What a skill is

A skill is a folder of instructions an agent reads when a task calls for it, in the Agent Skills format that Claude Code, Codex, OpenCode, Cursor, Gemini CLI and GitHub Copilot share. The agent sees only the skill’s one-paragraph description until a task matches it, then reads the rest.

The Orbitron skill is built into the orbitron binary. Installing it needs no download, no account and no network, and the copy it writes describes exactly the version that wrote it. See Skills.

What Orbitron sends anywhere

Nothing. Orbitron has no network layer: it reads files and writes files. The desktop’s agent connection, when you turn it on, is a socket on your own machine that only your account can reach. An agent using Orbitron sends whatever the agent itself sends to its own model provider, which can include the output Orbitron prints. That is between you and your agent; check its data policy before pointing it at unpublished results.

Checking an agent’s work

An agent can be confidently wrong. The skill teaches it to check warnings in the JSON output, to state a format’s tier, and to report the trust grade that comes with a metal site’s oxidation state, but read the numbers it reports against the files they came from before you rely on them.