website/docs/agents.md, in the nav under "Going further", is the player's page.
It covers what an agent can do, connecting it from the launcher, copy-paste
config for Claude Code, Codex, Cursor and any MCP client, the game's AI Agent
page, a CLI quick reference, safety and No AI.
docs/agent-control.md is rewritten for developers:
- the surfaces, and why one MCP server covers the launcher and the game;
- the inventory, before and after, for the launcher and the game;
- No AI on every surface, and the launcher commands that stay;
- connecting from the launcher and the game;
- the design rules for a new command.
docs/launcher-commands.md lists every launcher command, its arguments and its
MCP tool. It is generated from the clap definition (cli_tools::markdown). The
launcher test the_launcher_reference_is_up_to_date fails when it is stale or a
tool is undocumented, like horizon-ctl's generated reference.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The MCP tools are generated from the CLI's clap definition (cli_tools.rs):
mods_search is horizon-launcher mods search, with the same arguments, and
a call runs that command with --json. So the mod browser, modpacks, hosted
servers, settings, game arguments, data collection, desktop entry and the
benchmark are all reachable, and new CLI commands become tools by
themselves. Destructive ones need confirm: true and carry destructiveHint.
The session keeps play, server (with server_command/log/stop),
account_list and account_sign_in in-process, renamed after the CLI
(launch_client is play; play takes join for an invite code).
New CLI commands: settings [KEY VALUE] (selected version, folders,
install-new-versions, opening animation) and agent-config, the MCP config
for Claude Code, Codex, Cursor and any client. play takes --world and
--creative. Under No AI, mcp and agent-config are hidden and refused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>