Unreal from Codex: Unreal MCP setup, a session contract and a tested asset pipeline for coding agents
by PaperChaseWebb
Connect Codex, Claude Code, Grok Bot or Hermes to the Unreal MCP server in Unreal Engine 5.8 and keep every session in bounds: a session file with a checkpoint commit and a leave-out list, five risk tiers, owner approvals tied to file hashes, an action log with a verify result for every call, and recovery steps for editor restarts, stale tool lists, compiles and timeouts. Adds a generator to Blender to Unreal asset path with a headless Blender prep script, an import script that refuses changed files, a connection probe and a fifteen-check read-only checker.
- Price
- $24
- Difficulty
- Intermediate
- Setup
- 1 hr
- Format
- .md file
unreal-from-codex-mcp-session-contract-and-asset-pipeline.md · 62 KB · instant download after checkout
Want this built for you instead? We can build it for you.
What you end up with
Your agent works in a sandbox folder from a written objective, every write is checked and logged, nothing destructive or bulk runs without your yes, and every imported prop has a license, a triangle count and an approval for that exact file.
What you get
- 21 sections in one .md file
- Unreal MCP setup: plugins, starting the server, client configs and the write-once Codex entry
- Connection notes for Codex, Claude Code, Grok Bot and Hermes, keeping the server on your own machine
- mcp_probe.py: checks the server answers and lists its tools without calling any
- SESSION.md: objective, leave-out list, checkpoint commit, engine version and triangle budget
- Five risk tiers and an approvals ledger tied to file hashes, with a separate release yes
- An action log format with a verify result for every call, and recovery steps for restarts, stale tools, compiles and timeouts
- blender_prep.py: headless cleanup with joined parts, applied transforms, bottom-center origin, a triangle cap and a prep report
- ue_import.py: imports only approved, unchanged files into /Game/_Agent, never replaces, with a dry run
- ue_check.py: fifteen read-only checks, with a fixture builder and the expected output
- A rules block for any harness, ten prompts and a failure table
- A look at five Unreal MCP tools and plugins, with sources
Who it is for
Unreal 5.8 developers and technical artists who want a coding agent doing editor work and asset imports, and want to know exactly what it changed, with their yes before anything destructive or anything that leaves the machine.
Works with
Solid chips: the agent is named in this listing's own tools list. Dashed chip: Plain Markdown, so it also pastes into other agents such as Grok Bot, Hermes, OpenClaw or Muse. We list those as compatible by format only, not as tested.
Tools needed
Cost to run: The pack is the only purchase. Unreal Engine, Blender and Python are free to start; model usage and generators such as Tripo or Meshy are billed by those vendors.
Time to result: Connected, probed and through a first read-only session in about an hour
How it works
Test the check
Save the scripts from the file, run make_fixture.py and run ue_check.py on the three fixture folders.
Expected: 15 findings on seeded, 0 on fixed, 1 on broken-uproject; exit 2 on a missing folder.
Turn on the server
Commit the project, enable Model Context Protocol and All Toolsets, start the server and generate the client config for your agent.
Expected: python3 mcp_probe.py prints the server name and its tools, exit 0.
Write the session
Copy the Agent and Pipeline folders, create /Game/_Agent, and fill SESSION.md with the engine, the checkpoint commit, the objective and the leave-out list.
Expected: ue_check.py shows no U02 to U05 findings.
Run a read-only session, then one small edit
Add the rules block, use prompt 2 to map the level with tier 0 calls, then prompt 4 for one tier 1 edit you can see.
Expected: Every call has a row in actions.csv with a verify result, and the edit reads back as asked.
Bring in one prop
Run blender_prep.py on a model in Pipeline/raw, approve the FBX by its hash, run ue_import.py with --dry-run, then for real in the editor.
Expected: The asset is in /Game/_Agent/Imports, its bounds match size_cm, and ue_check.py prints 0 findings.
Sample prompt from the pack
Read Agent/SESSION.md. Using tier 0 calls only, list the actors in the level named in the objective, grouped by class, with their locations. Log each call in Agent/actions.csv with what you saw.
Prompt 2 of 10. Runs after the rules block in section 15.
I want to rename every texture in {folder} to the T_ prefix. Do not do it yet.
List the old and new names, the number of assets, the referencers you found and how to undo, then ask me.Prompt 5 of 10. A rename is tier 3, so it waits for your words.
Free preview of the file
# Unreal from Codex: Unreal MCP setup, a session contract and a tested asset pipeline for coding agents Version 1.0, 2026-10-28. By PaperChaseWebb. Works with Codex, Claude Code, Grok Bot and Hermes. Plain Markdown, five Python 3 scripts and a fixture builder, tested as described in section 14. The Blender step was run headless on Blender 5.2.3 LTS. The checker, the probe, the import dry run and the free sample checker were run on Linux against a seeded fixture and a mock MCP server. The Unreal Editor steps were not run in Unreal for this release, and each one is marked where it appears. Unreal Engine 5.8 puts a lot of power one sentence away from your agent. Its Unreal MCP plugin lets a coding agent list actors, change materials, spawn things, run editor Python and, with the wrong prompt, delete folders. This pack is how to connect Codex (or Claude Code, Grok Bot or Hermes) to that server and keep the session boring: a written objective, a commit before the agent touches anything, a log of every tool call with a verify result, and your yes before anything destructive, bulk, or anything that leaves your machine. The second half is the asset path most people try next: a model from Tripo, Meshy or your own Blender file goes through a headless Blender cleanup, gets a license row and a hash, waits for your approval, and only then is imported into a sandbox folder in your project. A read-only checker, ue_check.py, reads the project and tells you in fifteen checks where the sess
Worth knowing
- The Unreal MCP plugin is labeled experimental by Epic in Unreal Engine 5.8, and tool names can change between builds. List and describe the tools at the start of every session.
- The Unreal Editor steps were not run in Unreal for this release. They follow Epic's 5.8 documentation and a public Codex setup tested on 5.8.3 under Windows. Start with a read-only session.
- Keep the MCP server on 127.0.0.1. Do not open its port to your network or the internet to reach a remote agent.
- Read your generator's current terms before using its models commercially, and write them in the license column.
- On 2026-10-28, ue_check.py reported the fifteen expected findings on the seeded fixture, 0 on the fixed one and 1 on the broken project file, with file hashes unchanged and exit 2 on a missing folder; blender_prep.py ran headless on Blender 5.2.3, refused over the cap, decimated to 4,850 triangles when asked and refused to overwrite; the import dry run planned one import and refused a changed file; the probe passed against a mock MCP server. Run the tests on your own machine first.
Questions
Do I need Codex?
No. The setup covers Codex, Claude Code, Grok Bot and Hermes, and the rules block is the same text for all of them. Codex is in the name because that is where the Unreal skills work people are sharing started.
Can Grok Bot drive my editor?
Grok Bot's computer cannot reach the MCP server on your PC by itself, and it should not. Register your PC so Grok Bot can run the probe, checker and pipeline scripts there with your approval, or use Grok Bot as planner and reviewer while Codex or Claude Code makes the editor calls.
Why tie approvals to a hash?
So your yes covers the file you looked at. If a model is re-exported after you approve it, the hash changes, the import script refuses it and the checker flags it.
How is this different from Engine Agent Kits?
Engine Agent Kits sets Godot, Unity and Unreal up for an agent in general: rules files, build loops and editor scripting. This pack is the Unreal MCP layer on top: the server, the session limits, the approvals and the import gate.
Details
Version 1.0 · Updated Oct 28, 2026 · Markdown version of this page
License: Use on your own projects and on installs you do for clients and teams. Do not resell or republish the pack itself. Download terms
Related

The Unreal agent session contract and a five-check quick check (free sample)
From Unreal from Codex: the session file a coding agent reads before it touches your Unreal editor (checkpoint commit, objective, leave-out list, local endpoint), five risk tiers, an action log format with a verify result per call, five rules to paste, and a read-only quick check with a worked example.

Engine Agent Kits: Godot 4, Unity 6 and Unreal 5 rules files, editor bridges and build loops for coding agents
One kit per engine for coding agents: a rules file, a trap table of version mistakes, terminal commands to check, test and build without clicking, ways to give the agent eyes, asset pipelines with a license ledger and a one-milestone work loop.

Game Dev with AI Agents: method, prompts, skills and check loops
A Markdown playbook for building games with Claude Code, Codex or Claude: a seven-stage method, a staged brief, 16 prompts, five agent skills, self-check loops, engine notes and 30 game concepts.

Hermes Overnight Worker: Blender asset skills that run while you sleep, with the owner's yes at both ends
An unattended Blender worker for game props. The owner approves tonight's queue, the agent builds each prop headless inside triangle, minute and critique budgets, writes only to a staging folder, a separate critic session reviews every preview, and nothing moves into the real project until the owner says yes in the morning and a read-only check comes back clean. Real scripts, tested on Blender 5.2.3, with a seeded fixture of fifteen planted faults.

Consistent Asset Sets: one approved reference, style files and review sheets for AI-made art, props and cards
A small production system for making many AI-made assets that belong together: approve one reference, write STYLE.md from it, request only deltas, build props as code, and review every batch on one sheet with a ten-point check.

Game Jam Kit: a 48-hour Godot 4 plan and a two-day Unity PRD that a coding agent can follow
A tested jam method for working with a coding agent: seven milestones per engine, each with a first version, a first test and a written list of what to leave out, a JAM.md the agent cannot change without your new yes, three owner approvals (scope lock, art and audio licenses, upload), a scope-cut checklist and a fifteen-check read-only jam_check.py. Includes a playable Godot 4 starter that tests itself headless, and a tested proxy for games that call an LLM while people play.

Make a 3D product ad with Claude Code and headless Blender on an old GPU
A 15-second, four-shot product spot rendered on a 2016 graphics card. Claude Code wrote the whole Blender scene in Python, Blender rendered it with no window open, and ffmpeg put the shots, music and overlay together. The builder posted the render times and what broke.

Agent Readiness Self-Check: score one workflow before you hand it to an AI agent
A free self-check for agencies and small businesses. Score one workflow across eight areas (workflow mapped, data access, approval gates, owner per agent, handoffs, cost tracking, security and vetting, measurement) with an evidence rule and three gate rules, read your band and what to do next, plus an interview prompt for Grok Bot, Claude Code, Codex or Hermes, a score sheet and a worked example.

Game Assets as Code: one spec file, seeded generators and a fifteen-check checker for agent-made game art and sound
Let Grok Bot, Claude Code, Codex or Hermes make your placeholder and jam-ready game assets as code instead of prompts: pixel sprites from letter grids, seeded symmetric sprite variants, seamless tiles, SVG icons, retro sound effects and low-poly Blender props built headless, all from one spec file with a fixed palette and seeds. Every build writes a manifest with file and spec hashes and a review page; your approvals are tied to those hashes, a promote step copies only approved, unchanged files into the game, and a separate release yes covers anything that leaves your machine.