BELLATRIX for AI Coding Agents

A model-free automation CLI plus skills and subagents that teach your AI coding agent to write BELLATRIX tests for web, desktop and mobile.

What it is

The BELLATRIX AI agent toolkit has two parts:

  • The bellatrix CLI, a .NET global tool with one command per UI action: bellatrix desktop|web|mobile <verb>. It runs no AI model and holds no credentials. Every command prints one line of JSON.
  • Skills and subagents that bellatrix init writes into your repository, in the native layout of your AI coding agent. They teach the agent the BELLATRIX component, page-object and test APIs of your language, and how to drive the CLI.

Your AI agent does all the reasoning. It reads the skills, drives a live app through the CLI to find stable locators, then generates BELLATRIX page objects and tests in your project.

Both parts are open source under the Apache 2.0 license: BELLATRIX-Agent on GitHub.

How it works

  1. Scaffold once. bellatrix init writes the skills, a locator-specialist subagent and a test-generator agent per domain, plus one instruction fragment that your agent loads.
  2. Drive the app. The agent starts a session (start-session), captures the UI tree to a file (snapshot) and interacts (click, type-text, …). Sessions persist across separate CLI calls by name; there is no daemon.
  3. Keep the context small. A snapshot returns only a file reference. Locators are resolved by the isolated locator-specialist subagent, or by a narrow bellatrix <domain> find --snapshot query. The agent never reads a full UI tree into its main context.
  4. Generate code. With validated locators, the generator agent writes the page object (actions, map, assertions) and the test in your language.
bellatrix init --agent claude --domains desktop,web,mobile
# then ask your agent, for example:
#   "Write a BELLATRIX test that converts 1 meter to feet in the Windows Calculator."

The driver, browser server or device is yours to run. For web, the CLI starts and stops a local browser driver itself unless you point it at your own Selenium server.

Supported agents and languages

AI agents: Claude Code, GitHub Copilot, Cursor, Devin Desktop (formerly Windsurf), Codex CLI, Gemini CLI (enterprise — Gemini Code Assist), Antigravity CLI (agy), Kiro CLI and OpenCode.

Languages: C#, Java and TypeScript. C# and Java cover desktop, web and mobile (Android and iOS); TypeScript covers web.

Each language × domain × agent combination has its own verification status on the support matrix page. "Partial" means the files are placed in the agent's documented layout but have not yet been verified live in that tool.

Next steps