MCPBridge Server
# MCPBridge Server
Companion MCP server for the **MCP Bridge** Unreal Engine 4.27 editor plugin.
It translates [Model Context Protocol](https://modelcontextprotocol.io) tool
calls from AI coding agents (Claude Code, Codex, Cursor, or any MCP client)
into HTTP commands for the listener that the plugin runs inside the Unreal
editor.
```
AI agent (MCP client) --stdio--> this server --HTTP localhost:8080--> UE4.27 editor (MCP Bridge plugin)
```
All communication stays on your machine: the server talks only to
`localhost:8080`, where the plugin's in-editor listener accepts connections
from the local machine only.
## Requirements
- Node.js 18 or newer
- The MCP Bridge plugin installed and enabled in a UE 4.27 project
(available on Fab; the plugin's Quick Start covers editor setup)
## Setup
```bash
npm install
npm run build
```
Then register the server with your MCP client. For Claude Code, add this to
your project's `.mcp.json` (adjust the path to where you cloned this repo):
```json
{
"mcpServers": {
"unreal-bridge": {
"command": "node",
"args": ["<path-to-this-repo>/dist/index.js"]
}
}
}
```
Start your UE 4.27 project with the plugin enabled, then ask your agent to
run `test_connection`. A healthy reply includes the engine version and
project paths.
## What the tools cover
150+ tools across: actor and level editing, Blueprint creation and
compilation, Blueprint member/graph editing with schema-validated pin
connections, gameplay framework generation (GameMode/Character/controller/HUD
with class defaults wired), input mappings and control-scheme presets, camera
rig presets, Blackboards and Behavior Trees built from JSON, C++ class
generation and UnrealBuildTool compilation as background jobs with structured
error reporting, materials and material instances, DataTables, audio
components, map creation, viewport control and screenshots, and a searchable
project intelligence index.
Every editor mutation runs in a UE4 transaction (undo works), compiles and
saves the affected asset, and reports failure explicitly - a broken Blueprint
is never reported as success.
## Tests
```bash
npm test # unit tests against a mock listener; no Unreal needed
npm run test:integration # requires UE 4.27 running with the plugin listener
```
The registry-consistency test cross-checks this server's tools against the
plugin's Python command router; it runs in the plugin's development
repository and skips automatically here.
## Versioning
Server and plugin versions track each other (this is v0.4.0, matching plugin
v0.4.0). When updating one, update the other.
## License
MIT - see [LICENSE](LICENSE).
TDQS
Scored across 122 tools
Several tools overlap: blueprint_info, blueprint_inspect, and blueprint_document all inspect Blueprint structure; blueprint_build_from_json and blueprint_build_from_description both build graphs; bridge_clear_log and clear_output_log are exact aliases. While many tools are domain-specific, these overlaps create selection ambiguity.
Most tools follow a consistent <domain>_<action> pattern (actor_spawn, asset_list, material_create), but outliers like undo, redo, help, test_connection, python_proxy, and ue_logs break the pattern. There are also mixed conventions such as level_actors vs actor_spawn and clear_output_log as a bare verb.
With 122 tools, this is an extreme over-scoping for a single server. The calibration guide marks 50+ as extreme, and this is more than double that, making it difficult for an agent to efficiently select the right tool.
The tool surface is remarkably broad, covering assets, actors, blueprints, materials, viewport, input, AI, data tables, and C++, but it lacks basic delete operations for many asset types (e.g., asset_delete, material_delete, blueprint_delete) and has no rename or list for data tables. It's extensive but not fully comprehensive.