// documentation

From zero to a playable game in minutes.

Install the plugin, point your agent at one local endpoint, and start prompting. Conduit turns natural language into levels, Blueprints, UMG, input, and a play-tested game — inside the real Unreal Editor. No sidecar, no glue code.

Start free — it's free for indies Jump to install

5 steps · ~10 minutes · Claude Code, Claude Desktop & any MCP client

65
typed, safety-gated tools
4
agentic skills that ship outcomes
0
sidecars — native, in-engine
// before you start

What you need

Unreal Engine 5.7 or 5.8 (broader 5.5–5.8 in progress)
A C++ project with the Visual Studio toolchain (Windows)
An MCP-capable agent — Claude Code, Claude Desktop, or similar

Python Editor Script Plugin and Enhanced Input are auto-enabled as plugin dependencies — nothing to install by hand.

// install

Install in five steps

It's a drop-in source plugin. Place it, build it, flip the switches you want, connect, and prompt.

1
Get the plugin

Grab a license (free for indies), accept your GitHub invite, and download AgentConduit-vX.Y.Z.zip from the repository's Releases tab. Unzip the AgentConduit/ folder.

Get Conduit free
2
Drop it in & build

Place the folder at YourProject/Plugins/AgentConduit/, regenerate project files, and build the Development Editor target.

# your project layout
YourProject/
  YourProject.uproject
  Source/
  Plugins/
    AgentConduit/        <- drop the folder here
      AgentConduit.uplugin
3
Enable & launch

Open the editor, enable Edit → Plugins → AgentConduit, and restart. On startup the Output Log confirms the server is live:

AgentConduit MCP server listening on http://127.0.0.1:30010/mcp
4
Flip the safety gates you want

Conduit is read-only until you opt in. In Project Settings → Plugins → AgentConduit, enable only what you need — each is an independent switch:

Allow Editor Mutationswrite tools (spawn, author, edit)
Allow Console Commandsexec console + CVars
Allow Python Executionthe Python escape hatch
5
Connect your agent & prompt

Point any MCP client at the endpoint — full details in the next section. Then just talk to it.

// connect your agent

One endpoint. Any MCP client.

Conduit speaks the Model Context Protocol over Streamable-HTTP. If your agent talks MCP, it works.

Claude Code — one command
claude mcp add --transport http agentconduit http://127.0.0.1:30010/mcp
Claude Desktop / config-based clients
{
  "mcpServers": {
    "agentconduit": {
      "type": "http",
      "url":  "http://127.0.0.1:30010/mcp"
    }
  }
}

Locked-down setup? Set an Auth Token in settings and every request must carry Authorization: Bearer <token>. Empty by default for open local connections.

// your first prompt

Say it. Watch it build.

Describe what you want. The agent picks the tools, makes the edits inside the real editor, and verifies the result with screenshots and play-in-editor.

You > Make a coin-collector: a floor, a coin Blueprint with a Score
      variable and a BeginPlay graph, spawn one, a GameMode, and
      SpaceBar mapped to jump.
// Conduit drives the editor
New level created and saved, floor actor with a real cube mesh
BP_Coin with a StaticMesh component and a Score variable
Event graph: BeginPlay → PrintString → SET Score=5 → Branch, compiled clean
Coin spawned, BP_GameMode applied, Enhanced Input SpaceBar → IA_Jump

A verified live run — 24 of 24 tool calls succeeded against Unreal Engine 5.7.

// skills

Outcomes, not API calls

Conduit ships expert Skills — workflows that encode design judgment. Install them into your project and your agent gets a head start on real game-building.

pwsh Scripts/Install-AgentConduitSkills.ps1 -ProjectRoot "C:\Path\To\YourUnrealProject"
game-bootstrapper

A whole playable loop — level, pawn, input, GameMode, HUD, win/lose — proven in PIE.

gameplay-logic-builder

A wired, compiling Blueprint graph from a plain behavior description.

game-ui-builder

A designed UMG interface — HUD or menu — with layout, type scale, and styling.

level-designer

A blocked-out, lit, dressed scene with real composition.

// the toolbox

65 typed tools, grouped by job

Every tool has a real schema, validation, and an undo transaction. The full, always-current reference ships in the repo as docs/Tools.md.

World & level
new/load/save level, game mode
Actors & scene
spawn, transform, pose, scatter, components, physics
Blueprints & logic
create, variables, components, node graphs, connect pins, compile
UMG — game UI
widget blueprints, add widget, properties, render-to-image
Materials & look
materials, params, decals, post-process
Data & assets
create/import/save assets, properties, arrays, curve tables
Play & verify
camera, PIE, screenshots, reference diff, output log
Batch & escape hatches
batch calls, Python, console, CVars (gated)
// safety

Trusted in a real project

Read-only by default. The agent can inspect everything but change nothing until you opt in.
Transactional. Every write runs in an undo transaction — Ctrl-Z works, and failures roll back.
Serialized. Structural edits are queued so overlapping requests can't corrupt an asset.

Your next game starts with a sentence.

Free for indie developers. Install in minutes, prompt your first level today, and ship when you're ready.