Monolith
One plugin. Every Unreal domain. Zero dependencies.
Why I built this
Most MCP integrations for Unreal register every action as a separate tool. That floods the AI’s context window with hundreds of tool names before you’ve asked a single question — and the actually useful stuff gets buried. I built Monolith because I wanted my AI to spend its context on my problem, not on memorising a tool catalogue.
One plugin. One MCP endpoint. A handful of namespace-dispatch tools instead of ~1,400+. The AI calls monolith_discover() and monolith_guide() when it needs to know what’s available, and otherwise just hits blueprint_query("create_blueprint", ...), material_query("compile", ...), and so on. monolith_discover() is terse by default — each action returns just a name and a one-line description, so listing a namespace costs almost nothing; the full param schema is one describe_query("action_schema", ...) call (or detail=true) away when the AI actually needs it.
I use it every day. It does what I need.
What it does
Monolith exposes ~1,400+ actions across 26 in-tree namespaces through a namespace-dispatch pattern: each domain registers a single {namespace}_query(action, params) tool, and a central monolith_discover() lists everything available. (Exact counts are intentionally approximate — query monolith_discover() for the live figure.)
Covered domains: Blueprints, Materials, Animation, Niagara, Mesh, UI (incl. CommonUI), AI (Behavior Trees, State Trees, EQS, Smart Objects, Perception, Navigation), Gameplay Ability System, Logic Driver state machines, ComboGraph combo trees, Audio (Sound Cues + MetaSounds), Editor control (UBT builds, log capture, scene capture, asset preview & inspection), Engine source search (1M+ symbols, fully offline), Project asset search (SQLite FTS5), INI config, Level Sequences, a bulk_fill / describe reflection framework for deep property writes, a monolith_guide self-onboarding tool for your AI, plus the new v0.17.0 Reflection Intelligence layer: decision (architectural decision-record harvest), risk (repo-level hotspot + co-change + conditional-gate signals), cppreflect (UE 5.7 UHT reflection-edge queries cross-joined with the asset registry), network (replication inspection — replicated classes, RPCs, OnRep handlers, unbalanced-handler audits), pipeline (read-only composer actions for PR review + release pre-flight), and reflect (index maintenance — a project-only force-rebuild of the reflection tables). The cppreflect and network indexers scan your project plugins by default, so replicated classes and RPCs declared in plugins are in scope without extra setup; enabled marketplace plugins are gated behind a setting, and Epic engine built-ins stay excluded.
MCP LLM Ergonomics (also new in v0.17.0): universal response shaping (_fields / _omit / _compact_json) on every action, schema-tagged param kinds with automatic \ → / rewrite on asset paths, did_you_mean fuzzy match on dispatch errors, MCP tools/list annotations (read-only / destructive / idempotent hints), source_query cursor pagination, and a proxy-side JSONL call log. The whole point is to let your AI spend less context recovering from typos and trial-and-error.
New in v0.19.0: an LLM C++ authoring ergonomics pack in the source namespace — eight read-only lookups so your AI resolves an include path, exact signature, deprecation status, Build.cs deps, header lint, or a UCLASS stub in one round-trip instead of reading raw source (get_include_path, get_signature, check_deprecations, verify_symbols, find_example_usage, suggest_build_cs_deps, lint_header, generate_class_stub), plus fix_hints on editor.get_build_errors. A parser fix that finally indexes allman-brace plain classes/structs tripled the engine source index (~300K → ~967K symbols), so lookups for FCollisionShape, FScopeLock, FPaths and ~40K other engine types now actually return. Plus live-PIE introspection + driving in the editor namespace (pie_get_object_properties, pie_call_function, pie_set_control_rotation, pie_inject_input_action, pie_possess_spectator_free), programmatic stat-group readout (get_stat_group_values), time-series PIE sampling and anim-node binding read/write (animation), a Blueprint variable-reference census (blueprint.find_variable_references) and contract reconciliation, and first-class T3D asset-text export (project.export_asset_text). The tools/list manifest is ~40% smaller (duplicated action lists dropped from dispatcher descriptions), and two first-launch fixes land (MonolithMesh now delay-loads GeometryScripting; the deep indexer no longer asserts on UserDefinedStruct fields with unresolved types — issue #70, thanks @aggitti).
New in v0.20.0: an AnimGraph-authoring pack in the animation namespace — apply-additive / mesh-space-additive nodes, slot nodes (validated against the skeleton’s slot groups), save/use cached pose, output-pose and state-result wiring, blend-by-int, sync groups, layered-blend-per-bone filters, Control Rig anim-graph nodes, linked anim layers, and state-machine conduits — plus blend-space baking + interpolation control, state-machine teardown (remove states / transitions / re-point entry), IK-solver removal, and a Blueprint-Assist-free auto_layout formatter that works in release builds.
New in v0.18.1: a from-scratch Motion Matching authoring pack across the animation, chooser, and blueprint namespaces — Pose Search schema / database primitives, mirror data tables, chooser-table authoring, the AnimBP motion-matching graph + foot-IK, thread-safe AnimBP authoring (reflective Property Access, a thread-safe function flag, and an exec-driven chooser feeding the Motion Matching database), character/actor scaffolding, and a retarget create/run pack. Plus a PIE / profiling harness (async PIE-smoke sessions, CSV / Insights profiling brackets, clip + anim-frame capture, map authoring, nav rebuild/validate), state-machine authoring + live anim-instance telemetry, a generic AI controller that runs a BehaviorTree on possess with movement-driving BT task classes, inherited-native-component inspection, and live DataAsset field read-back.
New in v0.18.0: Niagara HLSL direct-editing — read and overwrite the HLSL source on a CustomHlsl node (get_custom_hlsl_text / set_custom_hlsl_text), plus simulation-stage / event-handler selectors on the module-stack actions and a ParameterMap bridge for create_module_from_hlsl (PR #65, thanks @middle233). Niagara also gains a search & discovery pack (search_by_parameter, search_by_data_interface, query_niagara, find_similar_systems, search_by_material, find_niagara_references, list_system_data_interfaces).
Full per-namespace breakdown: Tool Reference (wiki).
Works with Claude Code, Cursor, Cline, or any MCP-compatible client. Windows, macOS, Linux.
Quick install
1. Drop into Plugins/
cd YourProject/Plugins
git clone https://github.com/tumourlove/monolith.git Monolith
(Or grab the latest release zip and extract to the same path. The release zip includes precompiled DLLs so Blueprint-only projects can open the editor immediately without rebuilding. Monolith builds on UE 5.7 and 5.8 from a single source tree — but the precompiled DLLs are engine-locked, so Blueprint-only users grab the zip for their engine, Monolith-vX.Y.Z-UE5.7.zip or -UE5.8.zip. Building from source works on either.)
2. Create .mcp.json in your project root (same directory as your .uproject):
{
"mcpServers": {
"monolith": {
"command": "Plugins/Monolith/Binaries/monolith_proxy.exe",
"args": []
}
}
}
The native C++ proxy keeps your AI session alive when the editor restarts. For Cursor/Cline, macOS/Linux, or the Python fallback, see the Installation wiki page.
3. Open the editor. Wait 30-60 seconds for the first-launch index. When you see Monolith MCP server listening on port 9316 in the Output Log (filter LogMonolith), connect your AI client and ask “what Monolith tools do you have?” to verify.
Project-instructions files (CLAUDE.md, AGENTS.md, .cursorrules, etc.) vary per assistant — just paste the namespace list into your AI and ask it to generate the right format for your toolchain. Full install variants, troubleshooting, and post-install setup live on the Installation wiki.
Two zero-dependency C++ executables ship in Binaries/ and work without the editor:
monolith_proxy.exe — MCP stdio↔HTTP proxy. Keeps your AI session alive across editor restarts. Used by the .mcp.json config above.monolith_query.exe — Offline query tool. Serves the engine source index, project asset index, and the full Reflection Intelligence surface (decision / risk / cppreflect / network) without launching UE — byte-identical to the live server, verified by a ship-blocking parity guard. Instant startup; useful for terminal-side lookups and CI when the editor is down.
Details: wiki Tool Reference.
Auto-updater
Off by default as of v0.14.6. Opt in via Auto Update Enabled in Editor Preferences > Plugins > Monolith — checks GitHub Releases on editor startup, verifies the downloaded zip’s SHA256 against the release-notes marker, swaps the plugin on editor exit (after a Y/N prompt). See Auto-Updater wiki.
Network exposure
Monolith starts a local HTTP server on port 9316 to receive MCP traffic. UE’s FHttpServerModule does not expose a bind-address parameter, so the listener is reachable on all network interfaces, not just 127.0.0.1. CORS is restricted to localhost origins (which blocks browser-based cross-origin reads) but does not block direct HTTP requests from other devices on the same LAN.
If you work on an untrusted network: either add a Windows Firewall rule blocking inbound TCP on port 9316 from non-loopback addresses, or untick MCP Server Enabled in Editor Preferences > Plugins > Monolith and restart the editor.
See SECURITY.md for the full threat model and disclosure policy.
Documentation
- Wiki — installation variants, tool reference, connecting your AI, configuration, auto-updater, FAQ, skills, optional modules, engine source index details, mesh module deep dive, horror level design, procedural geometry, genre presets, test status
- API_REFERENCE.md — full per-action parameter reference, regenerated from the live registry each release
- SPEC_CORE.md — technical specification and architecture; per-module specs at
Docs/specs/
- CHANGELOG.md — version history, contributor credits, breaking-change notes
- CONTRIBUTING.md — dev setup, coding conventions, how to add new actions, PR process
Contributing
Contributions welcome. See CONTRIBUTING.md. Every release CHANGELOG names the PR authors and issue reporters whose work shipped — credit goes where it’s due.
License
MIT — see ATTRIBUTION.md for credits.
评论