GumCli for Agents

GumCli (gumcli) is a cross-platform command-line tool for working with Gum projects without opening the editor. Because it is scriptable, has machine-readable output, and reports meaningful exit codes, it is the primary way an AI agent acts on a Gum project and verifies its own work.

This page is a recipe-oriented tour from an agent-workflow angle. For the full reference on each command, follow the links into the CLI section.

Installation

gumcli installs as a .NET global tool:

dotnet tool install -g GumCli

Why it works well for agents

• Machine-readable output — command results (JSON, file paths, summaries) go to stdout, while banners and progress chatter go to stderr. An agent can consume stdout directly, for example gumcli check MyProject.gumx --json.
• Meaningful exit codes — 0 for success, 1 for errors found, 2 for a project that could not be loaded or invalid arguments. An agent can branch on the exit code instead of parsing prose.
• No editor required — everything runs headless, so it fits an automated edit → validate → verify loop.

The agent loop

A typical agent workflow for building or modifying a Gum UI follows three stages.

1. Generate code

If the project uses code generation, initialize the settings once, then generate strongly-typed C# for the project's screens and components. This eliminates magic strings and lets the agent reference UI elements by name.

• codegen-init — write the code-generation settings into the project.
• codegen — generate C# for the project's elements. Pass --prune to also delete generated files left behind by elements that were deleted or renamed.

2. Validate

After editing project XML — whether by hand, with custom tooling, or via an assistant — validate it before trusting it. check catches the silent mistakes that are easy for an agent to make, such as misspelled element names that Gum would otherwise drop without error.

• check — validate a project and report errors. A non-zero exit code tells the agent to stop and fix the problem before continuing.

3. Verify visually

Generated code that compiles is not proof the UI looks right. Rendering the result to an image gives the agent something concrete to inspect — and, with a multimodal model, to actually "see."

• screenshot — render a Screen or Component to a PNG. An agent can open the PNG to confirm the layout matches intent.
• diff-screenshots — render every Screen and Component via both monogame and raylib and report any mismatch, so a raylib-specific rendering bug doesn't slip through if the project targets both.
• svg — render to SVG when a vector format is preferred.

Supporting commands

• new — scaffold a new Gum project.
• fonts — generate missing bitmap font files referenced by the project.
• pack — bundle a project into a single .gumpkg file.

Pair gumcli with the MCP Documentation Server so your assistant can look up how a command or property works while it drives the project, and with the AI Skills so it understands the file formats it is editing.