Run an existing python script from the repo scripts/ folder through a headless commandlet. Optional args reach the script via sys.argv (no whitespace in an arg). Output is filtered to marker lines (default: 'MCP |LogPython') — have your script unreal.log() what matters.

READ-ONLY: load a map and read actor properties by their editor display names (e.g. 'Use Auroras', 'Ultra Dynamic Weather'). Matches actors whose label OR class contains actorMatch (case-insensitive). This is the tool shape that found the dead aurora switch and the null sky->weather reference — read back the ACTUAL state instead of trusting push logs.

Set editor-display-name properties on matching actors in a map, then SAVE the map. Values are converted to python literals safely (booleans included). Mutating: refuses while an editor is open unless allowWithEditorOpen.

Duplicate a /Game/... map asset. Mutating: refuses while an editor is open unless allowWithEditorOpen.

Assign a material to a slot on a named actor's static mesh component in one map, then save. Mutating: refuses while an editor is open unless allowWithEditorOpen.

READ-ONLY: report the sky (UDS) and weather (UDW) actors in a map with the fields that have actually burned us: aurora/space master switches, the sky->UDW reference, snow/rain state, manual-override pins, time of day. Returns one MCP INSPECT_JSON line.

Apply a named editor-time UDS lookdev preset to the sky actor and save the map. WARNING: runtime phase profiles (DA_TOD_*) override these every tick in-game — this is for in-editor look checks only. Mutating: refuses while an editor is open unless allowWithEditorOpen.

Server/environment status: paths, whether an Unreal editor is currently running (file-lock risk for mutating tools), and generated-script housekeeping.

READ-ONLY: query the asset registry under a /Game path, optionally filtering by class-name and asset-name substrings. Cheap discovery before probe/set tools.

READ-ONLY: read properties by display name off ANY asset — DataAssets (e.g. TOD profiles), meshes, materials — or a Blueprint's class defaults (classDefaults: true). The asset-side sibling of unreal_probe_actor.

Set simple-valued properties (number/bool/string; a /Game/... string is auto-loaded as an object reference) on an asset or a Blueprint's class defaults, then save the asset. Struct/array properties are NOT supported — use unreal_run_python_script for those.

Batch-spawn actors into a map and save. Each entry: source (a /Game static mesh or Blueprint asset, or a /Script/Module.Class), label, location, optional rotation/scale/tags. clearTag first deletes every actor carrying that tag (idempotent re-runs). NOTE: headless commandlets have no physics scene — no ground traces; supply explicit Z.

Batch-import files (FBX meshes, PNG/JPG textures, WAV) into /Game destinations via AssetImportTask. textureType configures compression: 'normal' (TC_Normalmap, sRGB off), 'mask' (sRGB off), 'color' (default). Files must live inside the repo. Big texture batches can exhaust headless memory — chunk them.

Create (or update) a MaterialInstanceConstant from a parent material and set scalar/vector/texture parameters, then save.

Import an animation-only FBX onto an EXISTING skeleton (the Trekker pipeline: no mesh, exported-time length, no default sample rate). Post-tune the sequence with unreal_set_asset_properties if needed.

Regenerate the level snapshot markdown (label, class, transform, tags for every actor) that CLAUDE.md declares the authoritative source of level truth. Writes inside the repo.

Run a repo script, then — in the SAME commandlet boot — probe an actor or asset and check expectations against the read-back values. The whole call fails unless the script succeeds AND every expectation passes. 'Probe before you believe', as infrastructure: a mutation that can't prove itself didn't happen.

THE EYES: launch a -game instance of a map, wait for it to settle, teleport through vantages (BugItGo) and/or run console commands, screen-capture each, kill the game, return PNG paths. This is the pattern every visual verification this project ever did was built on. Windows-only; steals foreground focus while capturing. Serialized with all other runs.

One-call arc verification: set a TOD profile's driver to a fast FixedTimer, launch the map in -game, detect the run start from the log, capture a screenshot as each PHASE becomes active (named by phase), kill the game, and restore the original driver. Needs the machine idle-ish: captures are focus-verified and report NOFOCUS if you're typing elsewhere.

READ-ONLY animation audit for a skeleton: every AnimSequence (play length, rate scale, additive), BlendSpace (sample positions + which sequence at which speed), Montage, and AnimBlueprint under a search path, plus which assets the AnimBP actually references (orphan detection via the asset registry). The state-machine GRAPH itself is Blueprint territory — audit that via the live UnrealClaude plugin, not headless.