Druids AI - Druids MCP Bridge

Druids MCP Bridge gives MCP-compatible AI assistants a structured way to inspect, create, edit, and refactor projects directly inside Unreal Editor.

Use your own AI account, model, or local LLM. Druids does not provide the AI, require a Druids subscription, or receive your prompts and project data. The bridge runs locally between your configured AI client and Unreal Editor.

Connect an assistant through the guided Connect AI Assistant window, then work naturally:

  • Inspect, search, create, and refactor complex Blueprints.

  • Audit and complete Blueprint Interface implementations from trusted project patterns.

  • Create and populate Primary Data Assets and structured properties.

  • Inspect and author supported Material graph operations.

  • Build and adjust deterministic PCG forests and city layouts.

  • Place actors, modify transforms, and capture viewport screenshots for visual review.

  • Search project assets, graphs, comments, variables, functions, materials, StateTrees, and supported specialized assets.

  • Preview complex changes, require approval, compile, verify, save, or roll back safely.

Druids MCP Bridge includes guided setup, structured capability discovery, signed update support, and detailed failure reporting so agents can clearly distinguish supported operations from genuine boundaries.

Requirements

  • Unreal Engine 5.5–5.8

  • Windows 64-bit

  • Unreal Editor use only

  • An MCP-compatible AI assistant or local agent

  • Optional Unreal plugins are required for their related features, such as PCG

The AI assistant is not included. Your AI provider’s own terms and data practices still apply.

Documentation

1 Like

We have a lot of updates and plan in the works. But let me know if there is a specific feature set you want to see sooner than later.

2026-8-21a

Features

  • Learn Blueprint presentation conventions from trusted project examples, including execution direction, spacing, variable placement, and comment styling, then use those conventions to analyze and improve selected graph regions or an explicitly approved whole graph.
  • Review a complete Blueprint presentation plan before applying it, with exact node moves, comment changes, task annotations, risks, and before/after metrics; approved changes compile, verify, and save together or roll back together.
  • Add task markers and notes directly to regular Blueprint nodes without replacing the text of surrounding comment boxes, and capture framed, opaque before/after graph images for visual review.

Bug Fixes

  • Blueprint cleanup now uses full node geometry and deterministic collision resolution so wide nodes, branches, shared data providers, and variable getters do not overlap after layout.
  • Selected-region cleanup preserves unrelated graph areas, existing comment text and style, node annotations, pin defaults, and every execution or data connection.
  • Presentation requests now distinguish regular-node annotations from comment-box editing, and weak or conflicting style evidence is reported as a fallback instead of being treated as a proven project convention.
  • Blueprint presentation and graph capture workflows now build and run consistently across Unreal Engine 5.5 through 5.8.

Coming in the Next update!

2026-8-29b

Features

- Create a protected optimization lab from an existing map, with separate immutable baseline and working copies that can be reviewed, reset, or removed through an approval-gated workflow without modifying the source content.

- Audit an exact map for rendering optimization candidates across texture memory and streaming settings, material complexity, mesh LODs and Nanite state, collision, repeated components, cull distances, lights and shadows, foliage, instancing, World Partition, and HLOD evidence. Results rank useful profiling targets while clearly separating static risk indicators from measured runtime cost.

- Capture repeatable rendering baselines as three locked 2560x1440 Epic-quality PIE trials, with matched camera and workload evidence, per-trial screenshots, bounded Rendering traces, native reports, and median Frame, Game, Render, RHI, and GPU timings for trustworthy before-and-after comparisons.

- Diagnose whether a protected rendering scene is CPU-, GPU-, memory-, streaming-, or loading-bound by combining scene audits, comparable PIE captures, and bounded traces, then receive evidence-backed reversible experiments and exact asset candidates for further testing.

- Optimize selected Static Meshes inside a protected lab by enabling Nanite or increasing minimum LOD on duplicated assets, retargeting only the working map, and verifying the saved mesh settings and map references before asking for visual review and a comparable follow-up profile.

- Test approved component-level rendering changes inside a protected working map, including shadow casting, light draw distance and attenuation, primitive cull distance, and collision mode. Changes are previewed as an exact plan, saved transactionally, captured for visual review, and kept separate from optimization claims until a comparable follow-up profile supports them.

- Find compatible groups of repeated Static Mesh actors and convert an explicitly approved group to ISM or HISM inside a protected working map. Plans preserve every source transform, mesh, material, mobility, collision, shadow setting, tag, and folder; verify the saved reduction in actors and primitive components; capture the visual result; and require a comparable profile before claiming a performance improvement.

- Run a complete protected-scene optimization workflow from map cloning and rendering audit through matched profiling, diagnosis, one reversible experiment, one approved persistent candidate, visual review, and a local evidence dashboard, while keeping the original source content unchanged.

- Open an exact saved map directly through Druids while choosing whether unsaved map changes should block the switch, save first, be discarded, or cancel the request. Saving and discarding require explicit approval, while unrelated dirty content assets remain untouched.

- Review optimization work in a result-first local dashboard that explains what was tested, identifies the exact persistent change, compares before-and-after screenshots and metrics, shows the evidence chain, and recommends one trustworthy next action.

Bug Fixes

- Failed lab creation, reset, or removal now restores or removes only the exact lab packages involved, preserving the original map and avoiding unsafe world-undo behavior during rollback.

- Rendering comparisons now reject mismatched resolution, quality settings, RHI or shader model, camera viewpoint, runtime scenario, workload, or capture protocol while preserving the measurements as diagnostic evidence. Temporary viewport, camera, and quality changes restore automatically, and captured PNGs use an opaque alpha channel for reliable viewing.

- Rendering diagnosis remains explicitly inconclusive when measured evidence is missing or conflicting, and static asset risks are never presented as proven runtime causes without a controlled experiment.

- Static Mesh optimization plans preserve source assets, materials, collision, sockets, and unrelated map references; unsupported meshes, stale plans, incomplete World Partition references, failed rebuilds, or verification errors stop safely and roll back the derived asset and working-map changes together.

- Instancing plans reject incomplete actor groups, external actor references, custom behavior, incompatible settings, World Partition maps, and stale evidence before replacement; failures before or after saving restore the exact original actors and package state without leaving a partial aggregate actor.

- Protected-scene verification now follows the exact selected audit, diagnosis, profile, trace, experiment, and persistent-change evidence into the local dashboard, and withholds an optimization claim when a visually valid candidate is slower or the measurements do not support an improvement.

- Performance dashboards now distinguish measured improvements, regressions, negligible changes, invalid comparisons, inconclusive evidence, pending visual review, and candidates that still need profiling. The verdict is derived from saved evidence rather than agent-written summary text.

- Render-change verification now persists exact asset-state, source-protection, visual-review, and profile-pair evidence, so an editor restart cannot silently associate a persistent change with unrelated captures and instead clearly requests a fresh comparable profile when required.

- Generated optimization maps now release their off-screen world and physics resources without invalidating saved map actors still referenced by the editor, preventing both world-leak and delayed Typed Element cleanup crashes. Rendering audits also expose canonical component settings that match plan and apply results for reliable post-save verification.

- Map switches now report the exact dirty map or external-actor packages involved, preserve One File Per Actor changes correctly when saving, and refuse invalid maps, active PIE sessions, unresolved editor dialogs, or untitled Save As decisions instead of risking hidden data loss.

- Runtime dashboards no longer display unavailable Render or RHI timing as zero cost. Profiles now report each timing source and its reliability, the dashboard shows `N/A` when evidence is not trustworthy, and rendering diagnosis excludes those values while directing deeper investigation to the bounded Rendering trace.

- Project Health now distinguishes unavailable inventory evidence from a measured zero, reports the complete count of ranked rendering leads, and explains each bounded audit gap directly in the dashboard Evidence view.