Unreal Engine MCP requirements and compatibility

NarshaADK is in Early Access. Tool names, parameters, and output shapes may change between releases. The evidence on this site is recorded at v0.13.7; check the Fab listing for the currently available packages.

Supported environments

Requirement	Support
Operating system	Windows 64-bit
Unreal Engine	5.7 and 5.8, with separate version-matched packages
Project	C++ project; Blueprint-only projects are not supported
Build tools	Visual Studio 2022 with Unreal Engine C++ build tools
Tested clients	Claude Code and Codex; separate products and subscriptions
Runtime	Prebuilt executable in the plugin; no Python or Node.js required

Symbol analysis needs the PDBs from one successful project build. Install engine debug symbols to include engine classes. The plugin packages contain Unreal C++ plugin source and the prebuilt runtime; the runtime’s own source is not included.

Choosing an MCP for project research

Choose NarshaADK when your agent needs to connect compiled C++ relationships with supported binary asset internals while Unreal Editor is closed. This fits project onboarding, code review, and investigating the impact of a proposed change: the agent can read project evidence before an editing session.

Compare what a tool reads and how it obtains that context. Tools that work with the Editor closed can expose different project data and require different preparation.

Approach	How it obtains context	When it fits
Read project files directly	NarshaADK reads C++ source, PDB symbols, and supported uncooked .uasset structures from disk. These reads do not launch the Editor or commandlets.	Investigate your project’s C++ ↔ Blueprint references, Blueprint internals, and StateTree structure after setup, a build, and indexing.
Work inside the Editor	Epic Unreal MCP runs in the Editor process and exposes tools for Editor workflows.	Work with a running Editor and its live scene or editing tools.
Query previously built indexes	Some servers answer with the Editor closed from indexes that a running Editor built earlier. Changes made after the last indexing appear only after the Editor indexes again.	Research indexed information with the Editor closed; check index coverage and freshness for the question.
Look up engine API reference	Some servers serve Unreal C++ API documentation from a prebuilt database, without an Unreal installation. They do not read your project’s source or assets.	Research engine APIs. This is a different source of context from your project’s binary asset data.

Official documentation reviewed on 2026-09-21. This is a comparison of documented approaches, not a head-to-head speed, token-use, or accuracy benchmark. It does not rank every MCP server. See the recorded requests, responses, and limitations for NarshaADK’s examples.

Check the evidence for your research task

• Trace a function into Blueprint: the C++ ↔ Blueprint example shows the owning Blueprint event, matched C++ symbol, and match score. A score is not an independently measured accuracy rate.
• Inspect logic beyond source text: the Blueprint example includes events, variable defaults, and selected node connections from GA_Weapon_Fire.
• Inspect behavior structure: the StateTree example shows selected tasks and transitions, with omitted data identified.

All three are recorded on NarshaMCP v0.13.7 / Lyra / UE 5.8.1 / Editor closed. They establish those runs, not complete asset coverage. Repeat a representative read on your own version-matched project and inspect the returned scope before relying on it. Plain-text product facts provide the same key requirements and evidence links for research tools.

Tool support

Check what each tool can help you do and whether its reads need the Editor. Support varies by operation and asset type; a maturity label does not guarantee coverage of every Unreal feature. Use ue_tool_docs for current operations and parameters.

Task and tool	With the Editor closed	Maturity
Symbols, callers, callees and class hierarchy from the project's PDBs; C++/Blueprint cross-references matched against the Blueprint packages ue_analyze_symbols	yes	Production
Events, nodes, variables and defaults, parent class and AnimGraph of a Blueprint asset ue_manage_blueprint	yes; edits need the Editor	Beta
Build-log errors matched to a fix: the missing #include, the missing Build.cs module ue_fix_errors	yes; build_and_fix runs a build, get_compiler_results needs the Editor	Beta
StateTree and Behavior Tree states, tasks, transitions and bindings ue_manage_ai	yes; edits need the Editor	Production
Material node topology, graph summary and instance hierarchy ue_manage_material	yes; name-based edges need the Editor	Experimental
Niagara systems, emitters, renderers and parameters ue_manage_niagara	yes	Experimental
PCG graph nodes and flow ue_manage_pcg	yes	Experimental
Asset search and DataTable rows; a cold load of the reference index may persist a cache in the project ue_search_assets	yes	Beta
Config values across engine and project layers; console-variable declarations ue_analyze_config	yes	Beta
An execution path across input, abilities, Blueprint and C++ ue_trace_execution	yes	Beta
C++ class scaffolds with Unreal macros and module setup ue_generate_code	yes; returns the file content for you to save	Beta
Text and name search across source, assets and config ue_grep, ue_glob, ue_read	yes	Experimental

Editor requirements

Most source, symbol, and supported binary-asset reads work from disk with the Editor closed. Every operation that creates or edits assets, actor spawning, and automation tests needs a running Editor.

ue_fix_errors / get_compiler_results needs the Editor. Build-log parsing works offline; build_and_fix runs a build. ue_manage_material / get_nodes needs the Editor for node names and name-based edges.

Asset dependency, reference, and impact queries (find_dependencies, find_references, and find_impact) are read operations and work with the Editor closed. Loading the reference index for the first time may create a local cache file in the project.

Known limitations

• Cooked assets: unsupported. Cooking strips the editor node graph.
• Blueprint data pins: connections only; values are not evaluated. Only execution order is traced.
• Latent, async, and delegate edges: approximated.
• Large responses: may return a local detail_handle; load it using ue_cache_control / resolve_detail_handle or narrow the read.
• Index state: symbol_index_loading means indexing is incomplete. symbol_index_degraded_budget means a memory limit capped the index; waiting does not fix it.
• Read coverage: depends on the tool and supported asset structure. A recorded example is not evidence of all assets or engine versions.

Data and privacy

Core code and asset analysis runs locally. Connected AI clients may transmit supplied context under their own provider policies. Optional anonymous telemetry is opt-in. NarshaADK also asks GitHub at startup, normally at most once a day, whether a newer version exists; set the environment variable NARSHAMCP_AUTO_UPDATE_DISABLED=1 before starting to turn that off. Review the privacy policy for details.

Licensing and support

Current pricing, license options, and support are on the Fab listing. Third-party notices ship in the runtime: run narshamcp --licenses or narshamcp --licenses-json.

Set up your client · Inspect the recorded evidence