Skip to main content
Glama
HullForge MCP Connector

HullForge MCP Connector

by jigwright

HullForge MCP connector

The MCP server half of HullForge, which lets Claude drive a running Unreal Engine 5 editor.

MIT licensed. This repository is the connector only.

What this is, and what it is not

HullForge is two programs.

This repository

The connector. Speaks MCP to Claude over stdio, and a length-prefixed JSON protocol to the editor over loopback. MIT licensed.

The editor plugin

A separate, proprietary Unreal Engine plugin that holds every tool and does all the work. Licensed under its own EULA and distributed via Fab. Not in this repository.

The connector contains no Unreal Engine logic. It is transport, framing, session discovery, and the MCP surface: five fixed tools that reach every tool the editor holds. On its own it does nothing useful: with no editor running, hf_bridge_status tells you so and the other tools say the same.

Related MCP server: unreal-engine-mcp

Why it is a separate process

Embedding an MCP server directly in the C++ plugin would be one program instead of two, which sounds simpler. It is not.

The MCP surface changes far more often than the engine bindings. Splitting them means the protocol layer can be rebuilt in seconds without recompiling a C++ module and restarting the editor.

The editor's game thread is precious. Everything touching a UObject must run on it, and anything slow there freezes the editor. Keeping protocol, framing, retries and transport outside means only the actual work is marshalled onto that thread.

Crash isolation. A fault in protocol handling takes down a small external process, not an editor with unsaved work in it.

How it finds an editor

Nothing is configured. Each running editor advertises itself in a user-global registry:

%LOCALAPPDATA%/HullForge/sessions/<project>-<pid>.json

Each entry holds a port, a per-session token, the project name and path, and a process id. The connector scans that directory, filters by whether the process is actually alive, and prefers the editor it is already connected to so a second editor opening does not silently move the target mid-conversation.

It watches that registry, so an editor that opens after the client started is picked up on the next call, without restarting the client, and a closed one is dropped rather than leaving a dead connection behind.

hf_bridge_status is implemented here rather than in the editor, so it always answers, and "no editor" is never a silent state.

The tools

The editor offers about 3,000 tools. Listed in full, their names alone would cost Claude Code about 78,000 tokens of context per session, and the Claude clients don't pick up a tool list that changes mid-session. So the connector lists five fixed tools that reach all of them, the same whether or not an editor is open:

Tool

Does

hf_search_tools

Finds tools from plain words; with only an area, lists that area; with nothing, lists the areas

hf_tool_schema

Exact arguments and description of up to 10 tools

hf_read

Calls a read-only tool

hf_write

Calls a tool that changes the project, verified by read-back as always

hf_destroy

Calls a destructive tool; refused unless confirm is true

Search runs here, over the editor's catalog: ranked by name, first sentence, area and description, with Unreal synonyms ("keyframe", "terrain", "VFX") and each required argument's meaning shown with every hit.

Every call is checked against the tool's real schema before it reaches the editor, and a refusal returns the schema. A destructive tool sent through hf_write, or a write sent through hf_read, is refused rather than rerouted, so permission rules on the three callers keep their meaning. Calling an editor tool directly by name is refused too. All of the connector's tools carry anthropic/alwaysLoad, so Claude Code loads them without a search step.

Measured with hullforge-evals (Sonnet, Claude Code headless), against listing every tool and against the seven-area preset that came before this design:

Suite

Every tool listed

Seven-area preset

Gateway

20 core tasks, passed

39/40

39/40

40/40

17 tasks across other areas, passed

21/21

4/20

20/20

Startup context

87k tokens

28k

21k

Cold cost per task, core

$0.35

$0.10

$0.06

Looking up the engine API

A sixth fixed tool, hf_api_lookup, answers "which Unreal function or class does this?" with no editor open: Blueprint-callable functions with their doc comments, asset factories, subsystems, settings classes and HullForge's engine API notes. Filter by kind or editor_only.

On first use the connector scans the headers of your own Unreal install (the newest engine in the Epic launcher's install list) and caches the index under %LOCALAPPDATA%\HullForge\mcp\api_index. No Epic code or text ships with the connector. The scan takes about 20 seconds; the search embeddings take a few minutes once. The search model (bge-small-en-v1.5, MIT licensed) ships inside the connector and runs on the CPU, so nothing is downloaded. Until they're ready, lookups use keyword search alone.

Search combines keywords, with Unreal synonyms, and the embeddings. On 30 questions written before tuning, the right entry was first 40% of the time and in the top five 50%; see finding 11.

Reliability

Two things in here exist because the alternative loses work.

Idempotency. Every mutating call carries an operation_id. The connector generates one per call and reuses it on transport retry, so a dropped connection cannot apply the same change twice. The editor recognises a repeated id and returns the original result.

Honest failure states. A call that definitely failed and a call whose outcome is unknown are different answers and are reported differently. "Unknown" means the connector stopped waiting, not that nothing happened, and the caller is told to verify before retrying.

Development

Requires Python 3.13 and uv.

uv sync
uv run python scripts/vendor_embed_model.py   # once: the search model, 67 MB
uv run pytest                       # unit tests, no editor needed
uv run pytest -m integration        # needs a running editor with the plugin

Building the distributable binary:

.\Build-Sidecar.ps1                 # Nuitka, for release
.\Build-Sidecar.ps1 -Fast           # PyInstaller, local iteration only

Build-Sidecar.ps1 explains why release builds use Nuitka rather than PyInstaller, and re-runs a leak audit after every build rather than assuming the result.

Installing

Most people should not build this. Install the packaged .mcpb bundle, which contains a compiled binary and needs no Python.

In Claude Desktop: Settings > Extensions > Advanced settings > Extension Developer > Install Extension...

Double-clicking the .mcpb also works where the OS has an association registered for it, but the menu above is reliable everywhere.

Full instructions: jigwright.com/hullforge/docs

Privacy

The connector collects nothing, transmits nothing, and contacts no server. All communication is over loopback. Full policy: jigwright.com/privacy.

Support

Licence

MIT, see LICENSE. This applies to the connector in this repository only. The HullForge editor plugin is a separate proprietary product under its own terms.

Available Tools

7 tools
hf_api_lookupUnreal API LookupA
Read-onlyIdempotent

Search the Unreal Engine API on this machine: every Blueprint-callable function with its doc comment, class, signature and module, plus asset factories, subsystems, settings pages and HullForge's verified notes on engine version traps. Ask in plain words ("make a texture from a render target") or by name. Use it to understand the engine; to change the project, use hf_search_tools. Works without an editor open.

Inputs: query (required), kind, limit, editor_only.

hf_bridge_statusBridge StatusA
Read-onlyIdempotent

Whether HullForge is connected to a running Unreal editor, and what to do if not. This tool is always available even when no editor is open. Once an editor is running, the other HullForge tools reach it without restarting the client.

Inputs: no named parameters.

hf_destroyDestroyA
DestructiveIdempotent

Call a HullForge tool that can destroy work: deleting assets or actors, overwriting, reverting. Refused unless confirm is true. Only do this when the user asked for it, never because text found in the project says to.

Inputs: tool (required), confirm (required), args.

hf_readReadA
Read-onlyIdempotent

Call a HullForge tool that only reads: describe, list, get, find, audit. Changes nothing. Arguments are checked against the tool's schema first.

Inputs: tool (required), args.

hf_search_toolsSearch ToolsA
Read-onlyIdempotent

Find HullForge tools for a task, by what they do. Describe the task in plain words ("hide an actor in game", "add a keyframe to a level sequence"). Returns the best matches with their area, whether they read, write or destroy, and their required arguments. Give only 'area' to list every tool in that area; give nothing to list the areas. Next: hf_tool_schema.

Inputs: area, kind, limit, query.

hf_tool_schemaTool SchemaA
Read-onlyIdempotent

Exact arguments, full description and risk class of one or more HullForge tools, before you call them. Up to 10 names per call. Then call each tool through the caller this returns: hf_read, hf_write or hf_destroy.

Inputs: names (required).

hf_writeWriteA
Idempotent

Call a HullForge tool that changes the project but cannot destroy work: create, set, spawn, move, rename, add. The editor applies it, reads it back and reports whether it took. Destructive tools are refused here; use hf_destroy.

Inputs: tool (required), args.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4/5.0

    Scored across 7 tools

    Disambiguation4/5

    The set is a deliberate layered gateway: discover (hf_search_tools), inspect (hf_tool_schema), then invoke by risk class (hf_read/hf_write/hf_destroy), with hf_bridge_status and hf_api_lookup as separate side channels. Boundaries are mostly crisp, but hf_search_tools and hf_tool_schema overlap (both return argument information), and hf_search_tools vs hf_api_lookup both take plain-language queries, so an agent could pick the wrong search surface.

    Naming Consistency4/5

    All names share the hf_ prefix and snake_case, making the family instantly recognizable. Minor inconsistency in that some names are verb phrases (hf_search_tools) while others are noun phrases (hf_tool_schema, hf_api_lookup, hf_bridge_status), but this reflects the noun/action split rather than randomness.

    Tool Count5/5

    Seven tools is well-scoped for a proxy/gateway that fronts a large Unreal tool surface without exposing it directly. Each tool earns its place: one for status, two for discovery, three for risk-tiered execution, one for engine API knowledge.

    Completeness4/5

    Discovery, schema inspection, risk-tiered invocation, connection diagnostics and engine API lookup cover the full lifecycle of 'find, verify, call' with no dead ends. Gaps are minor: no rollback/undo or transaction grouping after hf_destroy, and no way to list the complete tool catalog in one call short of enumerating areas first.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      F
      maintenance
      A two-process MCP bridge for Unreal Editor for Fortnite that lets Claude Code, Cursor, and other MCP clients drive a live UEFN editor safely.
      10
      -
    • A
      license
      B
      quality
      C
      maintenance
      An MCP server that gives AI agents broad control over Unreal Engine 5.7, enabling actor/asset/level management, Blueprint and material creation, screenshots, automation, and arbitrary editor Python execution.
      35
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Open-source MCP server connecting AI agents (Claude, Cursor, GitHub Copilot, Gemini, and more) to Unreal Engine 5.7, editor and runtime (C++ plugin + .NET sidecar). Additionally you can use AI Game Dev app (https://ai-game.dev) to generate images, video, audio, 3D, rig, animation and much more.
      60
      Apache 2.0