Frequently Asked Questions
Which AI clients are compatible?
ClaudusBridge uses the Model Context Protocol (MCP), an open standard. It works with:
- Claude Code (CLI)
- Codex CLI (with the strict rmcp transport — fully supported as of 0.5.0)
- Claude Desktop
- ChatGPT (via MCP plugins)
- Cursor IDE
- Windsurf
- Any MCP-compatible client
You are not locked into any single AI provider.
Does it require a subscription or API keys?
No. ClaudusBridge is a one-time purchase. It runs 100% locally on your machine — no cloud services, no API keys, no recurring fees. Your AI client may have its own costs, but ClaudusBridge itself has zero ongoing charges.
Which Unreal Engine versions are supported?
ClaudusBridge requires Unreal Engine 5.7, supports development on Windows, Linux, Mac, and targets Windows, Mac, Linux.
Is the full C++ source code included?
Yes. You get the complete source code for the entire plugin — all 40+ manager classes, the MCP server, the router, the dashboard UI, the per-session event queue, the always-on internal watcher, the Neural Cortex, the auto-generated documentation pipeline, and all tool schemas. You can extend, modify, or learn from it freely.
What can ClaudusBridge actually do?
ClaudusBridge provides 492+ tools across 47+ categories:
- Blueprints: Create nodes, variables, functions, wire graphs, compile
- Materials: Create materials, expressions, instances, recompile
- Actors: Spawn, transform, configure properties and components
- Niagara: Build VFX systems, emitters, modules, renderers
- Widgets: Create UMG UI, set properties, bind events, animations
- Sequencer: Create cinematics, add tracks, control playback
- Procedural Mesh: Generate geometry, boolean operations, sculpting, fracture
- MetaSound: Build procedural audio graphs
- AI Vision: WebRTC
/previewstream + structured perception (ai_perceive,ai_vision_stream,get_vision,get_cognitive_map,scan_360,proximity_alert) - AI Navigation: Smooth camera movement, orbit, WASD-style navigation, fly_to_actor with collision avoidance
- AI Engineering: Survey, build, verify, audit construction workflows
- NNE: GPU inference via Tensor Cores for ONNX models
- Editor Events + Agent Bus:
subscribe_events,poll_editor_events,publish_agent_event,record_feedback— multi-agent coordination, self-improvement loop - And much more: Animation, Physics, Landscape, Audio, Navigation, Data Tables, MVVM, BlueprintLisp, Source Control...
How is it different from other AI editor plugins?
| Feature | ClaudusBridge | Typical Competitors |
|---|---|---|
| Protocol | Open MCP standard (spec 2025-03-26) | Proprietary |
| AI Client Lock-in | None — use any MCP client | Often locked to one |
| Tool Count | 492+ | 20–50 typical |
| Multi-Agent | Per-session events + agent_message bus | Single-agent only |
| Visual Verification | Built-in watcher closes the verification gap | Tool calls trusted blindly |
| Self-Improvement | record_feedback + feedback.jsonl accumulates plugin-improvement signal | None |
| Architecture | 100% local, no cloud | Often requires cloud |
| Source Code | Full C++ source included | Usually closed |
| Pricing | One-time purchase | Often subscription |
Is my data sent to the cloud?
No. ClaudusBridge communicates only over localhost. Your project data never leaves your machine through ClaudusBridge. The AI client you use may have its own data handling policies.
Can I extend ClaudusBridge with custom tools?
Yes. The architecture is designed for extensibility:
- Create a new Manager class (C++)
- Register commands in
ClaudusBridgeModule.cpp - Add tool metadata in
Resources/tool_schemas/ - Recompile — your new tools appear automatically
See the Connecting Your AI Client guide for the full developer flow.
Does it work in packaged/shipping builds?
No. ClaudusBridge is an Editor-only plugin (Module Type: Editor). It is not included in packaged builds and has zero impact on your game's runtime performance or shipping size.
What port does it use?
The MCP server runs on port 3000 by default. You can change this in Project Settings > Plugins > ClaudusBridge.
Can multiple AI clients connect simultaneously?
Yes — this is a first-class capability as of 0.5.0. Each connecting agent passes a unique session string in subscribe_events and poll_editor_events. Subscriptions are stored per-session, events are fanned out to every session subscribed to that type at insert time, and publish_agent_event enables agent-to-agent messaging on the same per-session queue.
Recommended convention: use the same string for clientInfo.name in initialize, session in subscribe / poll, and from_session in publish_agent_event so logs, dashboards, and event payloads all line up under one name.
This makes Claude Code, Codex CLI, Cursor, Windsurf, and any custom watcher / specialist agents able to work in parallel on the same Unreal session without stepping on each other. See Multi-Agent Coordination for swarm patterns.
What is the "verification gap" and how does ClaudusBridge close it?
MCP tools return status: success based on whether the API call returned normally. They do not verify that the visual / structural result actually matches the intent. Examples of silent failures:
spawn_actorreports success but the actor was placed inside a static mesh and is invisible.set_material_parameterreports success but the edited material instance isn't the one applied to the visible mesh.compile_blueprintreports success but emitted warnings that change runtime behavior.add_widgetreports success but the widget is at z-order 0 behind another widget.
ClaudusBridge closes this gap with three coordinated watchers:
- C++ Internal Watcher — Always-on, runs from plugin load. Inspects every editor event and fires three pattern rules (compile loops, PIE crash-on-start, actor leaks). Calls
record_feedbackdirectly. - Visual Watcher Sub-Agent — Mandatory canonical setup. Every connecting AI client spawns a sub-agent using the prompt at
Saved/ClaudusBridge/watcher_subagent.prompt.md. The sub-agent observes/preview?bare=1and reports observations on channelwatcher.observation. - Python Watcher Daemon — Optional persistent process for unattended verification. Pure stdlib, JSON logs, exponential-backoff reconnect.
When any watcher detects a discrepancy, it logs to Saved/ClaudusBridge/feedback.jsonl and broadcasts on the agent_message bus. The accumulating JSONL file is the canonical source of plugin-improvement signal across editor sessions.
What is Saved/ClaudusBridge/ and what files end up there?
Auto-generated on every MCP initialize (idempotent — files only regenerate when content changes):
| File | Purpose |
|---|---|
CLAUDE.md | Plugin architecture + full tool reference for AI assistants |
UE5_EDITOR_CONTROLS.md | UE5 keyboard / mouse shortcut reference for agents driving the editor through /preview |
watcher_subagent.prompt.md | Mandatory canonical sub-agent prompt for visual watching |
claudusbridge_watcher.py | Persistent watcher daemon (Python stdlib) |
claudusbridge_offline.py | Standalone MCP stdio server for offline tools |
Generated as plugin output; safe to delete (the plugin will rewrite them).
Accumulating data:
| File | Purpose |
|---|---|
feedback.jsonl | Plugin-improvement signal — append-only, written by record_feedback |
How do I report bugs or request features?
For bugs and feature requests, post on the Galidar Studio Discord.
For automatic plugin-improvement signal, your agents should call record_feedback whenever they notice a tool that didn't behave as expected, a tool they wished existed, a visual mismatch caught by the watcher, or any friction worth feeding back. The accumulating feedback.jsonl is reviewed regularly and turned into new tools, fixes, and schema improvements.