Your AI agent built a feature. Did it actually work? ProbeKit runs gameplay tests inside Unreal Editor and shows exactly what passed or failed — with steps, game-state checks and screenshots. No scripting required.
UE 5.7–5.8 · Windows · Editor-only · Needs an MCP AI agent · Single-player PIE tests
Your AI agent says "done". ProbeKit checks it in the game.
AI coding agents write Blueprint and C++ fast — but they never play the game, so "done" only means "it compiles". ProbeKit is an editor plugin that lets the agent play your game and check its own work.
What's included
the ProbeKit plugin, the ProbeKit panel in the editor and the built-in MCP server
- 12 example tests
- a sample actors module (editor-only)
- a PowerShell runner
- PDF documentation
What it does not do
It does not build game features by itself.
- It does not contain an AI model. You connect your own MCP agent (Claude Code, Cursor, Codex, others).
- It needs no ProbeKit subscription. Your agent may be paid.
- It does not judge art quality, visuals or game feel.
- It does not test several players at once (single-player play-in-editor only).
- It does not run headless: the PowerShell runner needs the Unreal Editor open.
How it works
You ask your agent for a feature: "Make the player respawn after falling into the lava." The agent makes it and writes a short test — a list of steps: walk into the lava, the player dies, the player comes back.
2. ProbeKit plays the test in the editor like a player: real keyboard input, real mouse clicks on your menus, live checks of the game.
3. The agent reads the report, fixes its own bug and tests again — until PASS.
Start in 3 steps: get ProbeKit and enable it; open the ProbeKit panel (toolbar button) and press Connect; write what should work, copy the request and paste it into your agent's chat.
Your first test. Ask your agent: "Run the ProbeKit example player_spawns and show me the result." It answers PASS with a screenshot of your level. The documentation also walks you through a first test of your own on a new Third Person template, using ProbeKit's sample spikes.
A panel, not a console. You do not need to write code or JSON. The ProbeKit panel in the editor connects your agent, turns "the door opens when the player walks up" into a request for the agent, runs every test with one button and shows the result in plain words: PASS or FAIL in colour, the step that broke, what the player saw. A failed test gives you a ready request for the agent to fix it.
The report the agent gets back:
PASS — every check held, with screenshots of what the player saw;
- FAIL — which step broke, and the game's own answer (the property value, the widget state, the log line);
- BLOCKED — the test could not run, so nobody mistakes "untested" for "works".
The tests stay. Each test is a small file in Tests/ProbeKit/. Weeks later, when another change breaks the same feature, the same test turns red — run it from the ProbeKit panel, the editor console or the PowerShell script, no AI needed.
Works with the agent you already have. ProbeKit is not another AI assistant. It has a built-in MCP server, so Claude Code, Cursor, Codex, Qwen Code, Cline, VS Code Copilot, Windsurf and any MCP client see its tools as soon as the editor is open — with a cloud model or a local one on your own machine. It sits happily next to other AI editor plugins. ProbeKit itself uses no AI and needs no subscription or API key; your agent keeps its own plan.
Nothing to install beside it. One plugin. No Node.js, no Python, no external server. The panel writes the config files for Claude Code and Cursor into your project and copies the setup for Codex or any other agent; restart the agent and it connects.
Safe by default. The server answers only on your own computer (127.0.0.1). An agent can play and check the game, but Python, command files, closing the editor and function calls outside the play session stay off until you tick "Let the agent use the whole editor" in the panel.
What a scenario can do
start and stop a solo play session (never touches a session you started yourself);
- run any console command — engine, cheats, your game's own;
- move the pawn through the game's real Enhanced Input actions (walk, jump, interact — no teleporting);
- click UMG widgets like a player, through a Slate virtual user — your own mouse never moves;
- read the on-screen widget tree: names, visibility, enabled state, rects, texts;
- read any property and call any Blueprint or C++ function on live objects;
- wait for a condition, check the log for a line (or its absence), take screenshots.
Tests (scenarios) are small JSON files; the agent writes them for you, and you can read and edit them. Commit them and anyone can rerun them — with an agent, from the editor console, or from the included PowerShell script while the editor is open (exit code 0 PASS, 1 FAIL, 2 BLOCKED).
12 ready examples on the stock Third Person and First Person templates: the player spawns, jump, a main menu's Play button, picking up an item, a door that opens, death and respawn, spikes that hurt, a checkpoint, coins, a locked door and key, a quest marker, a pause menu. Run them in the first two minutes, then copy them for your own features.
Try it free before you buy. Download ProbeKit Test Course, a small demo project with a free, limited build of ProbeKit already inside (precompiled, no C++ source; Windows; Unreal Engine 5.7 or 5.8): one level with four zones in a row (an item to pick up, a key and a locked door, lava with a checkpoint, a goal) and five ready tests. Unzip it, open it, press Run all my tests: the 4 gameplay tests pass, and the pause test needs the full edition because it clicks UI. After you buy ProbeKit, one button in its panel switches the demo to the full edition. UE 5.7: https://drive.google.com/file/d/187TosTMdVfhGypVMPAtJtNnMw_z1MER6/view UE 5.8: https://drive.google.com/file/d/1L58prYnmejv3dhf7F5eCththFADHuL8T/view
Free trial vs Full
MCP agent connection: Free yes, Full yes
- Real game input (Enhanced Input): Free yes, Full yes
- Read properties and UI: Free yes, Full yes
- PASS / FAIL with screenshots: Free yes, Full yes
- Steps per scenario: Free up to 8, Full no limit
- UI click, drag, wheel: Free no, Full yes
- PowerShell runner: Free yes, Full yes
The free trial is the same plugin with two limits, not a mock. It is a download from the link above, not a separate Fab product.
Safe by design. The server listens on 127.0.0.1 only and refuses browser requests. The modules are editor-only — nothing of ProbeKit is packaged into your game. Play settings are copied per run; your project settings are never changed.