RurixForge
中Download for Windows

RurixForge · Forge game worlds with AI.

Describe the game; agents plan, build, play-test and check it.

RurixForge logo

The workbench

One window for chat, plans, code and play

Sessions, chat, the workbench and the project pane side by side — plans, code and a demo you can play all open as tabs.

The RurixForge workbench: sessions, the conversation, a level script open in a tab, and the project pane

Real run · v0.1.0 + fixes · Oct 26, 2026

Recorded from a real run powered by Codex: UltraPlan is producing the 2D platformer Lantern Keeper, and the open tab is a level-data script the agents wrote. Compiling and passing checks doesn’t make it fun — that part is yours to judge. “+ fixes” means the capture used v0.1.0 with fix patches applied. The app’s interface is in Chinese; the OS user name and local paths have been replaced.

Modes

Eight modes. One composer.

Switch the mode in the composer and the agent switches how it works.

Agent01 / 08

Say it, and it gets to work

It reads the project, calls the tools and edits the scene — every step stays in the chat.

  • 127 always-on MCP tools
  • Three permission levels: Read-only / Approve writes / Auto-run
Real run · v0.1.0
Agent mode: the agent asks before each scene write, save and play step — all allowed here

Every screen is a real run, driven by Codex; the app’s interface is in Chinese. Output quality depends on the model, and passing automated checks doesn’t make a game fun.

UltraPlan

From one sentence to a demo you’ve played yourself

Describe the game in a sentence. It researches, asks a few questions and builds a playable demo for you to try. Only once you approve does it write the plan, staff a team and walk you through acceptance.

Real run · Lantern Keeper (re-run) · v0.1.0 + fixes · driven by Codex

  1. Step 1 of 7

    Discovery

    Say what you want to make; the agent reads your project and your brief to pin down the real ask.

    Real run · v0.1.0 + fixes

    UltraPlan discovery: the agent works out the gameplay and scope before writing the questionnaire
  2. Step 2 of 7

    Questionnaire

    This run: 23 questions, 15 handed to the agent in one click. Only the stack is yours to pick.

    Real run · v0.1.0 + fixes

    The submitted questionnaire: 23 questions, 15 of them left to the AI
  3. Step 3 of 7

    Web demo

    It builds an HTML version, then plays it in a real browser to check itself: 12/12 assertions passed.

    Real run · v0.1.0 + fixes

    UltraPlan building and verifying the playable web demo
  4. Step 4 of 7

    You play it

    The demo runs inside the workbench. Approve it or ask for changes — whether it’s fun is your call.

    Real run · v0.1.0 + fixes

    Demo review: the playable web demo runs right in the workbench; the formal plan only starts after it is approved
    Gameplay clip
    • ADmove
    • Spacejump
    • Elight

    Needs a keyboard — best on a computer. The game’s text is in Chinese.

  5. Step 5 of 7

    The real plan

    Only after the demo passes is the plan written: 13 tasks, 19 automated and 4 human checks.

    Real run · v0.1.0 + fixes

    The production plan: 13 tasks, 19 automated checks and 4 human checks
  6. Step 6 of 7

    Team production

    A team split by role takes the plan, up to 4 tasks at once; a failed check gets a targeted fix.

    Real run · v0.1.0 + fixes

    UltraPlan in production: agents split the plan, and the level-data script they wrote is open in the workbench
  7. Step 7 of 7

    Your acceptance

    After the checks, a few items are yours to verify by hand: feel, pacing, look. Passing checks doesn’t make it fun.

    AI illustration · this stage is not captured yet

Every captured screen and number comes from this one real run (Oct 25, 2026) and belongs to it — not to the product’s defaults. Output quality varies with the model and the brief.

Lantern Keeper web demo · made by UltraPlan in a real run

Design · Team

Built from the mockup, made by a team

  • Designfrom the mockup

    The title screen rebuilt in the engine, with editable text components
    The approved mockup

    Pick a 2D mockup; it splits the picture into elements and rebuilds them in the engine — titles and buttons stay editable text.

    Real run · v0.1.0

  • Teamside by side

    The UltraPlan production board: tasks in backlog, running, review and done columns

    One shared board, several agents working by role — 4 at once by default. You can step in at any time.

    Real run · v0.1.0

    The collaboration board of UltraPlan’s production stage

Showcase

Made in RurixForge. All of it.

A harbor built from a paragraph. A playable demo from one questionnaire. A mockup rebuilt, element by element, into a title screen.

A daytime harbor built by the agent: ivory sky, pale celadon sea, terracotta-roofed houses, a wooden pier, lanterns and a boat (Godot engine render)
Morning Harbor

AgentReal run · v0.1.0 + fixes

One paragraph to the agent: primitives, material graphs and a camera. When the colours came out wrong, one more sentence — it found the cause and fixed them.

Lantern Keeper · playable demo

UltraPlanReal run · v0.1.0 + fixes

The prototype UltraPlan builds right after its questionnaire: light a lantern, reveal the platforms in the fog.

Keyboard required
The same harbor from a low camera on the pier: terracotta lanterns and the houses on the shore (Godot engine render)
Morning Harbor · on the pier

AgentReal run · v0.1.0 + fixes

Same scene, one more request for a low camera — only the camera moved.

The title screen rebuilt in the engine, with editable text components
Mockup · rebuilt element by element

DesignReal run · v0.1.0

Design mode rebuilt the mockup in the engine; the title and buttons are live, editable text.

AgentReal run · v0.1.0 + fixes

One paragraph to the agent: primitives, material graphs and a camera. When the colours came out wrong, one more sentence — it found the cause and fixed them.

“+ fixes”: re-run on v0.1.0 after patching issues that stalled the production stage; those fixes are not in a release yet. What the AI produces varies with the model and the request — and passing automated checks doesn’t mean the game is fun.

Slide 1 of 4

Lantern Keeper · playable demo

Safety & control

You stay at the wheel

You decide what the agent may read and change: read-only, ask before every write, or let it run on its own.

  • Read-only

    In the app · 只读

    Can query, but can’t write to project files, scenes or assets.

  • Approve writes

    In the app · 写入需批准

    Reads run automatically; every write tool asks for approval first.

  • Auto-run

    In the app · 自动执行

    Reads and writes run automatically, still within tool and workspace boundaries.

    Default for new sessions

Force-deleting assets still in use, uninstalling packages, deleting skills: each raises a proposal and runs only once you approve it.

Real run · v0.1.0

Before it writes, it asks

  • 拒绝DenyThe step is skipped
  • 本会话允许Allow for sessionDon’t ask again this session
  • 允许AllowJust this once
Asked to recolor a title, the agent stops before writing to the scene and waits for your answer.
Real run · v0.1.0

After “Deny”

The change did not take effect … No shell command was run at any point.

— The agent’s reply, translated (excerpt)

Made a mistake? Step back

Scene edits undo step by step, and the built-in roles’ and skills’ playbooks have the agent save a scene checkpoint before big changes, so one step rolls them back. There are no automatic snapshots — keep the project’s history in your own git.

Under the hood

Inside the forge

Every step you see in the conversation is a concrete tool call underneath.

  • 8MCP servers
  • 127MCP tools
  • 10agent roles
  • 16built-in skills

¹ 127 is what the running v0.1.0 lists across its 7 always-on servers; connect Blender and blender-bridge adds 11 more.

One server per domain

Hover or tap a server to see what it does.

  • engine-scene60 tools · Scenes, entities and components, physics queries, play sessions and the viewport.
  • asset-pipeline27 tools · Asset import, builds, references and batch jobs.
  • store13 tools · The asset store: search, install and your personal library.
  • code-forge11 tools · Builds, diagnostics and structured edits for scripts and logic graphs.
  • gen-image7 tools · Image and texture generation.
  • context6 tools · Semantic index and search over your project.
  • gen-model3 tools · 3D model generation.
  • blender-bridge11 tools · Enabled once Blender is connected · Drives a running copy of Blender.

Two renderers, one workflow

Godot

4.7.2 · D3D12

The default for new 2D projects, and an option for 3D.

264/264combinations render

Godot’s own scene matrix: a render check, no pixel comparison.

Rurix

In-house · Vulkan · Jolt · Meshlets

The default for new 3D projects.

396/396frames identical

Regression baseline across the Godot integration, compared by frame hash.

ModelsNo bundled model: bring any OpenAI-compatible API, or sign in to Codex with ChatGPT.

The agent, the engine and your project files stay on your PC; only requests to the model and to the generation or store services you connect leave it. Output quality depends on the model and isn’t guaranteed.

FAQ

Questions, answered

Didn’t find your answer? Ask on GitHub

  1. Do I need to code, or bring my own model?

    No coding needed: say what you want and the agent plans, writes scripts and builds the scene while you steer. You do connect one model yourself: your own API key (OpenAI-compatible), Codex signed in with ChatGPT, or a cloud service you host.

  2. Which systems does it run on?

    Only Windows 10 (1809+) and Windows 11, x64 — no macOS or Linux build yet. The installer isn’t code-signed, so SmartScreen may step in: click More info, then Run anyway, and check the SHA-256 first if you like. Updates are manual: run the new installer over the old one; projects and settings stay.

  3. Is the quality of AI-made games guaranteed?

    No. Results depend on the model you connect — the same task can turn out very differently with another model, or on another day. Passing automated checks and playtest assertions means the game runs and meets its spec, not that it’s fun. The last test is yours to play.

  4. What else is limited in 0.1.0?

    Design mode recreates 2D screens only — hand 3D scenes to Agent or UltraPlan. The app’s interface is in Simplified Chinese only. There’s no one-click export for release yet: games run inside the workbench, and scenes, scripts and assets are plain files in your workspace.

  5. Where is my data? Is it open source?

    On your own computer — in %APPDATA%\RurixForge\data by default — with API keys encrypted by Windows DPAPI. RurixForge is open source under Apache-2.0; the code is on GitHub.

Didn’t find your answer? Ask on GitHub

RurixForge logo

What shall we build today?

Download for Windows

Windows 10/11 only for nowVersion 0.2.0 · free and open source

Source on GitHub