MythicLemon - Mythic Dev Assist | Agent Observability & MCP Bridge

Mythic Dev Assist — Agent Observability & MCP Bridge

Stop your AI agent from hallucinating in your editor.

Demo: Video

Setup Guide: Video | Documentation

User Guide: Documentation

Technical Reference: Documentation

You're using Claude Code, Cursor, or Codex to write your game. Your agent can write a Blueprint. It can refactor your C++. What it can't do is start your game, spawn the actor it just wrote, watch it tick, and read back what actually happened. So it guesses. And every guess compounds.

Mythic Dev Assist is the editor bridge that fixes this. It exposes 101 HTTP action routes and 14 standard Model Context Protocol tools so any modern AI coding agent can drive your UE5 editor — spawn actors, fire abilities, capture screenshots, sample Niagara state, run SQL against a 40-table session memory — and read the result back as structured, schema-stable JSON.

This isn't a chat panel. This isn't a Blueprint generator. This is the agent-native runtime the new generation of coding tools needs.

"Coding agents work well because the compiler acts as a source of grounded truth ... but in Unreal Engine that grounded truth is locked away behind UIs and binary file formats ... Mythic Dev Assist unlocks that so your agent can see what you see. It provided the necessary feedback loop for your agent to understand the world."

Other plugins give you 400 individual MCP tools that blow your context window up when the Python API can do 95% of the work anyway. MDA was build to solve real problems, not wrap them.

What you get

  • In-editor HTTP bridge (loopback only, default 127.0.0.1:7779, auto-cycles 7779–7788)

  • 101 action routes — spawn / destroy / move / call / compile / capture / import / preview / trace / query / open / close, all in one declarative table

  • 14 generic MCP tools wrapping the bridge: mda_observe_target, mda_search, mda_create, mda_set_target, mda_watch, mda_capture, mda_import, mda_query, mda_logs, and more

  • 18 always-on observability channels streaming live editor state into a per-session SQLite database — rendering, memory, GC, compilation, network, level streaming, scene hierarchy, animation, audio, AI, GAS, gameplay, niagara, physics, UI, entity tracking, log output, custom metrics

  • 40+ SQLite tables of session memory — your agent has both perception AND a queryable history

  • 4 Claude Code skills included — actions, capture, diagnose, query — for curated guidance inside Claude Code

  • Companion Python MCP server — one file, one dependency (pip install mcp), works with any MCP-compatible host

  • Dockable editor dashboard — live HTTP endpoint, session DB path, per-call action log with full request/response viewer

  • Local-first, zero vendor lock-in. Your agent calls localhost. No cloud account, no API key, no telemetry off your machine.

Why Mythic Dev Assist is different

Other "AI for Unreal" plugins sell the AI. Mythic Dev Assist sells what the AI sees. The plugin is built on 17 explicit agent-grounding design choices — what we call glass-box design:

World Pulse: Every response carries the live world state — frame number, PIE status, actor count, current map, timescale, game time. Your agent never needs a "where am I?" call between actions.

Project Twin: Engine version, plugins enabled, capabilities, target platforms, default map — attached to the first response of every session. Your agent stops guessing whether Niagara is enabled or whether the project is on UE 5.3 or 5.5.

Rich Mutation Responses: destroy returns the actual names and locations of the actors it removed. set_property returns the read-back value. move returns the post-collision-adjusted position. Your agent verifies its intent matched reality without a follow-up query.

Dry-Run / Preview Mode: Spawning, destroying, or executing console commands can run in preview mode — the agent sees the blast radius before committing.

Auto-PIE: Most agents forget to start Play-In-Editor. MDA auto-starts it when an action needs it and tags the response so the agent knows what happened.

Type-Aware Enrichment: Niagara responses auto-attach emitter counts, runtime particle counts, and active states. The pattern extends to materials, AI controllers, and other complex objects.

Niagara Preview Capture: One-call atomic VFX iteration — spawn, advance to N% intervals, screenshot each, return all frames with per-frame brightness, blank-ratio, and dominant-color analysis. No PIE needed.

Motion Telemetry & Kinematic Tracking: Entity tracking captures position, velocity, acceleration, jerk, and navmesh status on every snapshot. Find jittering NPCs with one SQL query: WHERE jerk > 10000.

Pattern-Based Proactive Warnings: After 3+ failures of the same action, MDA auto-attaches recovery guidance to the next response. The system teaches the agent to self-correct.

Semantic Agent Hints: Every error includes a specific recovery path — not "500 Internal Server Error" but "This requires PIE — start it with mda_world(action='sim_play')".

What you can build with it

Real workflows the included ue5-mda-* skills are tuned for:

  • "Spawn this Blueprint, set Health to 50, screenshot it from the front." One call each, structured response, paths returned ready for the agent to inline as images.

  • "Watch this Niagara system for 5 seconds and tell me whether it reached steady-state." mda_watch returns a per-sample timeline plus a delta summary plus screenshot paths.

  • "Find every actor on a steep slope right now." mda_query against the entity tracking table with a WHERE ground_slope < 0.7 clause.

  • "Why did this damage event do nothing?" mda_watch(trace_causality=True) links blueprint executions to property changes in the causal_executions table.

  • "What changed in the last 60 seconds?" Every channel has a timestamp. The agent writes SQL against the unified session DB.

  • "Capture a clean hero shot of this VFX for marketing." mda_capture isolate-mode produces lit, framed VFX shots with no editor clutter.

Technical highlights

  • Per-session SQLite database under <Project>/Saved/Observations/sessions/

  • PID-bound bridge identity — if your editor exits and another binds the same port, the next call is rejected rather than silently routed to the impostor

  • Auto-flush every 5 seconds; configurable

  • Editor-only — UncookedOnly modules are not linked into cooked shipping builds, so zero runtime cost in your packaged game

Works with

  • Claude Code (CLI + VSCode extension) — 4 curated skills included

  • Cursor — works via standard MCP config

  • Codex CLI — works via standard MCP config

  • Any MCP-compatible AI agent host

Requirements

  • Unreal Engine 5.3, 5.4, 5.5, 5.6, 5.7+

  • Python 3.10+ on your PATH (for the MCP server)

  • The mcp Python package — the only Python dependency

  • SQLiteCore, Niagara, GameplayAbilities — engine plugins, enabled automatically

Prerequisites

  • This plugin is the bridge layer between an AI agent and your editor. It does not include or require any specific AI provider — you supply your own Claude Code, Cursor, Codex CLI, or other MCP host.

  • No external network access. No cloud account required. Your project data stays on your machine.

  • First-time setup is ~5 minutes — install the plugin, register one MCP server entry in your AI client's config. Full walkthrough in the included User Guide PDF.

This plugin is actively used and developed, drop us a message on the forum with any feature requests.