Full installation

The quick start is the short path. This chapter adds everything it leaves out: the Unreal Engine skills plugin, upgrading an old install, the optional extras, what the first session does, and how to update a project without losing anything.

Two phases, two folders

Mixing the phases up is the most common source of confusion, so keep this table in mind.

Phase 1 in detail

The steps of the quick start (the .mcp.json file, Auto Start Server, copying agents and skills, registering the server) are Phase 1. These are the parts it does not repeat.

Check that the server listens

The default port is 8000. On Windows run netstat -ano | findstr :8000; on macOS or Linux run lsof -i :8000. A line that names the port as listening means it worked. No output at all means it did not: check that .mcp.json is in the right folder, that UEFN was restarted after the file was created, and that Auto Start Server is really on.

Upgrading from an older kit: delete the retired agents

Copying user-level-agents/ into ~/.claude/agents/ only adds or replaces files. An agent the kit has since renamed or retired stays on disk and can still be called by habit. In the console it shows up as an agent that is not in the roster. Before you copy a new version, compare the file list with what is in ~/.claude/agents/ and delete everything that is not one of the 17 agents listed in Your team.

Install Anthropic's Unreal Engine skills plugin

It is a separate plugin, maintained by Anthropic, and it complements this kit. It needs Git on your computer (Git for Windows on Windows).

Then run this inside Claude Code:

/plugin install unreal-engine-skills-for-claude-code@claude-plugins-official

Optional: a second brain in Obsidian

A separate wiki, outside any project, kept by the second-brain-librarian agent. Reusable mechanics and Verse implementations are written there once and found again in the next project. Skip this step and nothing else changes.

1. Create a vault anywhere on your computer (not inside a UEFN project) with the raw/, wiki/ and output/ folders, and copy second-brain-template/CLAUDE.md to the vault's root.
2. Check the file name from a terminal (dir CLAUDE* on Windows, ls CLAUDE* elsewhere). It must be exactly CLAUDE.md.
3. Open ~/.claude/CLAUDE.md and set the vault's full path as the "Second brain path" in rule 11.

Optional: shorter replies

The third-party caveman plugin makes Claude's replies much shorter and saves output tokens. It is not part of the kit and not required. The kit's own cost rules (cheap models, small/medium/large task sizing, history trimming) already do most of the saving.

Phase 2 in detail

Before the first session, in UEFN, open Project Settings for this project and enable Python Editor Scripting and UEFN MCP Toolsets. This is per project and cannot be skipped.

Copy the contents of project-template/ into the project's Content/ folder, open a terminal there, start claude and say: Read Claude/SETUP-INSTRUCTIONS.md and set up this project. The agent then:

1. reads the real project name from the folder above Content/ and writes it into CLAUDE.md;
2. checks that the MCP connection really answers from this folder;
3. asks you what the project is, its type and any conventions;
4. lists the available MCP tools and sets up the post-playtest hook;
5. checks that the project open in UEFN is the right one;
6. only then runs project-bootstrap for the first analysis.

You can leave SETUP-INSTRUCTIONS.md in place afterwards; it does no harm.

Updating a project

The easy way is described in Troubleshooting: with ~/.claude/kit-template/ in place, each session updates the kit files of the project you open. If you prefer to do it by hand, this is the rule.

Safe to overwrite at any time: Claude/hooks/ (the console, its server and every hook), Claude/reference/, Claude/community/, Claude/SETUP-INSTRUCTIONS.md, and .claude/settings.json (diff it first if you edited a hook matcher).

Never copy over: Claude/docs/ (SPEC, STATUS, ROADMAP, BUGS: the project's whole history), your CLAUDE.md, and Claude/logs/. They are the project's own data. The agents' per-project memory lives outside the folder entirely and is not touched either way.

After an update, hard-refresh the console tab. The old page may still be in the browser.

A project on a second computer

Registration of the MCP server is stored per computer, so a project you set up on computer A shows no MCP tools the first time you open it on computer B. That does not mean it needs a new setup. Do Phase 1 on computer B once; the agents register the server themselves when they notice it is missing.

Two files called settings.json

Your computer has two folders named .claude, each with its own settings.json, and they are never merged by hand.

Claude Code reads both and layers the project's on top of yours.