Ultimate Engine CoPilot - The World’s Most Comprehensive AI Development Tool for Unreal Engine

Distribution Product Discussion
84

How does one refund source code exactly? I am curious how that would work, just take your word for it that you deleted it? With this many questions and complaints without even attempting to use the trial yourself first, you sound like you will just be a problematic customer anyways and should just move on for everyone’s benefit. AI and this plugin are not a magic wand, you still have to know what you are doing in order to truly direct it.

2 Likes
85

Ultimate Engine Copilot这个工具我已经使用了3个多月了，虽然还是有些小的问题，但整体上确实超出了我的期望，希望2.0版本能尽快发布，能越来越好用！加油！

86

Literally says in the description and on the big pretty pictures that it uses local LLM aswell as gemini claude openai etc.

1 Like
87

Ultimate Engine Copilot V2 is Live!

Our Biggest update yet, full changelog can be seen here: Website Changelog

V2 Trailer

You can now try it out using a subscription if you don’t have the funds to purchase it right away: https://www.gamedevcore.com/pricing

Find it on FAB: https://www.fab.com/listings/8d776721-5da3-44ce-b7ef-be17a023be59
Or on the Website: https://www.gamedevcore.com/projects/bpgenerator-ultimate

88

V2.0.3, October 1, 2026

Fixes for the first reports on Version 2.

  • Games package again. With the plugin installed in the engine’s Plugins folder, where FAB and in-editor updates put it, packaging failed with “Missing precompiled manifest for UECPSteamRuntime”.
  • MCP clients stay connected when the editor restarts. Claude Code and other external clients no longer need a manual reconnect each time the editor is closed and reopened.
  • Custom endpoints such as Ollama now allow 16,384 output tokens by default instead of 4,096, so reasoning models have room to act. Custom Request Params written without the outer braces now apply, and the “response cut off” message shows the max_tokens that was actually sent.
  • Claude Code no longer declines every message. A declined message now starts a fresh session instead of sticking, and the plugin’s instructions reach Claude as a system prompt.
  • Long tasks keep going after compacting. Auto-compaction now runs just before your next message and then sends it, instead of stopping the chat after a paused or finished turn.
  • Your language sticks after a restart. The interface now starts in your language instead of English, and old per-project settings no longer overwrite it, or other settings, at launch.
  • Agent slots keep their model and effort, and installed agents show an Update button when a new version is out.
  • Fewer prompts in Auto Edit. Python asks only when a script deletes, renames or reverts something, and “Don’t ask again” now works.
  • MCP clients built on the Python SDK, such as Hermes, now see the tools that search_tools adds.
  • Import your own ComfyUI workflows in Local Studio.
89

V2.0.5, October 1, 2026

Fixes from your reports, including a detailed review of the MCP tools sent in by one of you.

  • Slot keybinds switch the chat you are in. Before, they only changed the default model, so a chat with its own model kept it.
  • BpLabs Standard works on installs that used Blueprints Lab 0.9. Every message there got “This model slot has been retired”. The old setting is now ignored.
  • Gemini with your own key no longer fails with a 400 error partway through a chat.
  • Tripo, Meshy and ComfyUI keys work whichever chat model is selected. A key entered while another model was active no longer goes missing.
  • Play test screenshots in chat include menus and HUD, and are saved to disk. With a dedicated server, input and screenshots go to the first client, and every reply says which world it used.
  • Runs beside SteamCore and other Steam plugins. CoPilot no longer turns on Unreal’s SteamShared plugin, and games that do not use Steam no longer ship its library.
  • Crash reports show where the editor crashed, and the crash prompt has “Don’t ask again” (also in Settings > Preferences).
  • DeepSeek and other custom endpoints no longer fail when an old “stream”: false is left in Custom Request Params.
  • Blueprint building is more reliable. Later builds keep nodes placed earlier, defaults set in a node entry are applied, enum pins take names or numbers, whole numbers stick on int pins, and a value a pin rejects is reported instead of dropped. Nodes stay where they were placed.
  • Project scans on large projects release what they load, stop before memory runs out, and no longer leave a healthy asset out of the index after one interrupted scan.
  • MCP: calls to localhost take milliseconds instead of about 2 seconds, a second editor on the same project shows its own address on the MCP screen, cooks no longer start MCP servers, and replies say when a dialog is blocking the editor.
  • Assets: new assets are saved when they are created (Material Instances too; Niagara systems you still save yourself), deleted assets are removed from disk, and FBX export runs without the options window.
  • Two changes for scripts: each Python call now starts with fresh variables, and property edits are no longer saved automatically (use save_asset).
  • Starting the editor no longer closes other apps that use ports 19878 to 19899.
90

V2.0.6, October 2, 2026

Fixes from your reports.

  • Crew has its three agents back. The Crew screen shows the Planner, Worker and Reviewer, each with its own model, what it is working on and a link to its chat, plus an activity log of every hand-off. The Planner is the chat you are in; pick a strong model there and cheaper ones for the Worker and Reviewer.
  • BpLabs Premium works for Crew agents. A Worker or Reviewer set to Premium, or following a Planner on Premium, ran on BpLabs Standard instead.
  • The roles you pick on a checkpoint are the ones that run. When the chat ran the plan, it could swap your Worker for Blueprint Dev.
  • You can stop a chat that carries on after a Crew result. The working indicator and Stop button were missing for that part of the turn.
  • Confirmation prompts show the whole call. Scripts show as code, every asset path the call touches is listed, and Show all opens the full text, so you can check exactly what will be deleted before you click Proceed.
  • Long editor operations no longer drop MCP connections. While a tool runs for minutes (heavy Geometry Script, a long Python script), the editor keeps the connection alive and reports progress, so your MCP client waits for the result instead of timing out.
  • Paste works on Linux. Ctrl+V and Shift+Insert in the CoPilot panel now paste from the system clipboard.
  • Cooks are left alone. CoPilot no longer edits DefaultEngine.ini while a cook runs. If an incremental cook misbehaves after you update CoPilot, run one clean cook.
91

V2.0.7, October 2, 2026

Fixes from your reports.

  • Chats can read your documents. A chat on BpLabs Standard or your own API key had no way to read a text file outside the project, so it asked for permission on every command-line read, or gave up. It now reads design documents, notes and data files (.md, .txt, .json, .csv and more) anywhere on your computer without asking. Files that hold keys or passwords stay off limits.
  • Command-line reads no longer ask every time. Commands that only look, like listing a folder or printing a file, run without a confirmation. Anything that changes files still asks first.
  • The AI Memory category picker works. Clicking it closed the list, and the editor with it, before you could choose. Memories waiting for review also no longer push your saved memories out of view.
  • The @ list shows where each asset lives. When several assets share a name, each one shows its folder, so you can pick the right one.
  • In-editor updates install on macOS. A Mac install could download an update but never apply it.
92

V2.0.8, October 4, 2026

Version 2 for more engines, and a round of work on local models.

  • Version 2 now runs on UE 5.5 and 5.6. Windows and Mac builds for UE 5.5, 5.6, 5.7 and 5.8. The MetaHuman extension needs UE 5.7 or newer.
  • Local models are no longer cut off after 30 seconds. A model on Ollama or LM Studio that took longer than that to read the prompt failed with “Connection failed”, every time. It now gets as long as it needs: Local Model Wait in Settings > AI Behaviour, 20 minutes by default. On Mac, restart the editor once after installing for this to take effect.
  • A chat stops itself when Blueprint graph builds keep failing. After 6 failed builds in a row it stops and says so, instead of looping until the conversation is full. The limit is a setting.
  • Blueprint tools understand more of what smaller models write. Graph instructions in the looser shapes weaker models produce are now read, and when something is missing the error names it.
  • Local Studio tells you what to expect. Use in Architect says whether your GPU has room for a larger context window, and a note says what small local models managed in our tests. See Local LLM Setup.
  • The MCP shared port can be changed. If another app already uses port 9876 (Blender’s MCP does), set a different one on the MCP screen.
  • Deleting one actor by its label works. The tool refused a single label and only took a list.
  • Fixed a crash on Mac when opening the CoPilot tab. It hit projects with a project index and a saved level open. UE 5.5 on Mac also no longer hangs or crashes under heavy MCP use.
93

V2.0.9, October 6, 2026

Agents that sign in with one click, a message on your phone when a task is done, and a round of fixes from your reports. Windows only for now: the Mac build stays on 2.0.8.

  • Sign in on the Agents tab works with one click. Press Sign in, a window opens, and your browser asks you to log in. No terminal and no separate install for Claude or Codex. It works for the other agents too, and a chat that fails because you are not signed in now tells you where to click.
  • Get a message when a task finishes. Paste a Discord or Slack webhook in Settings > Preferences and the plugin posts a short line when a long task is done, or when it is waiting for your answer. See Notifications.
  • GitHub Copilot’s answers show in the chat. Copilot finished its work and the reply never appeared. It does now.
  • Codex no longer burns tokens on a reviewer. Codex was started in a mode where a second model checked every step. It now asks you instead, through the plugin’s own confirm card, and that card says what it is asking.
  • Outside tools connect reliably through the shared MCP port. Claude Desktop, Claude Code and others could get “connection reset” on port 9876 while the editor was fine. Fixed.
  • Copy buttons tell you they copied. Every Copy button now shows “Copied”.
  • Links and dropped files no longer take over the panel. A file link in a reply shows the file in Explorer, and a file dropped on the chat is attached.
  • Replies follow your language. Write in Chinese with the interface in English and the answer comes back in Chinese.
  • “Sound when finished” plays again for BpLabs Standard, BpLabs Premium and your own API keys.
  • Fixed the most reported crash in 2.0.8. Opening a level from a chat could close the editor. Also fixed: a material that could no longer be opened after the AI changed one of its internal flags.
  • Fixed a long freeze on simple questions. Asking how many Blueprints a project has could lock the editor for a minute while every asset was loaded. It now answers in seconds.
  • Animation Blueprints: hidden pins can be exposed and bound. The AI can now do “Expose as pin” on any AnimGraph node, so a property like Reset Every Frame on Offset Root Bone can be driven by a variable.
  • Crash reports say what the editor was doing. A report now includes the last tools that ran, which is what we need to reproduce the crashes we could not fix yet.
  • The Learn leaderboard shows your account username. Some accounts showed another name, or “Anonymous”. It now always shows the username from your account page, or a placeholder like “Player 4821” until you pick one. It never uses your email address.
94

V2.1.5, October 9, 2026

A PaperZD extension, fixes for agents and local models, and a round of fixes from your reports. Windows only for now: the Mac build stays on 2.0.8.

  • New PaperZD extension. Build 2D character animation from a sprite sheet: flipbooks, animation sources, Animation Blueprints, state machines and playable characters, 39 tools in all. It is opt-in under Settings > Extensions and needs the free PaperZD plugin. See PaperZD.
  • Agents no longer stall when they edit several files. An agent that asked to edit two files at once could hang until you pressed Cancel. The requests now queue and appear one at a time.
  • Auto-compact works for local models. The context size your LM Studio or Ollama model was loaded with is now what the limit is measured against, older reasoning is trimmed as well, and a request that cannot fit is stopped with a clear message instead of crashing the server.
  • A notice when an update is waiting. A strip at the top of the CoPilot window offers Get update and then Restart now, so you no longer have to open Settings.
  • MetaHuman plugins no longer switch CoPilot off. CoPilot used to turn the MetaHuman Creator plugin on by itself, so disabling any MetaHuman plugin disabled CoPilot too. It no longer does: the MetaHuman extension turns that plugin on when you enable the extension. If you used the MetaHuman extension before, enable it once more under Settings > Extensions.
  • Typing with an IME works again. Text no longer vanishes while a Japanese or Chinese conversion is in progress.
  • Screenshots no longer overwrite each other. Every play-test screenshot gets its own file.
  • Variables can be marked Transient, like the checkbox in the Details panel.
  • Ctrl + mouse wheel changes the text size.
  • Long replies have a Copy button at the bottom as well as the top.
  • The question card is smaller on small screens, so the answer it refers to stays readable.
  • Updates are stricter. CoPilot never installs an older build over a newer one, drops an update left over from an earlier session, and a device you blocked on your order can no longer download updates.
  • Closing the editor after the AI ran a script is safer. The scripts’ leftover state is cleared before the editor shuts Python down.
95

V2.2.1, October 13, 2026

Fixes from your reports, and Azure OpenAI with the newest GPT models. Available on Windows and Mac.

  • Azure OpenAI and gpt-5 / gpt-6 work on custom endpoints. Requests to these models were refused because they use a different token setting, and gpt-6 only accepts tools through the Responses API. CoPilot now sends the right setting, uses the Responses API when your endpoint is a /responses URL, and switches to it by itself when the endpoint asks for it. A null in Custom Request Params removes that field from the request, and Test Connection checks the same request chat uses.
  • Saving a level with a Timeline no longer fails with “Illegal reference to private object”. New Timeline curves are created correctly, and Blueprints that were already affected are repaired the next time they compile.
  • Codex uses the model you picked. The model name is matched to the one the agent offers, together with its reasoning effort, for example gpt-5.6-terra with medium. If the agent still refuses a model, the task is stopped with a message instead of running on a different model.
  • A provider that answers with an error shows its message. Some providers, GLM among them, answer a wrong endpoint URL or auth header with a normal reply that contains an error, and CoPilot showed “The model ended its turn without a reply”. It now shows what the provider said, so you can see what to fix in the slot.
  • Indexing a project with Control Rig assets no longer crashes. The project scan leaves those assets alone.
  • Importing a material graph no longer crashes.
  • Niagara. Reading a running system no longer crashes on some custom engine builds, and binding a module input now compiles the system and reports the result.
  • Image generation providers. Choosing a provider in the list now fills in its endpoint and model, and Stability AI and Recraft are supported. Flux and Ideogram are removed from the list because they never worked there.
  • The Add slot button works again when some of your older slots were saved by an earlier version.
  • AI scripts no longer hold on to editor objects. Python scripts the AI runs no longer keep actors, assets and Blueprint classes between calls, which stops stale references piling up after a recompile. Numbers, text and functions you define still carry over.
96

V2.2.5, October 15, 2026

Voxel Plugin support, a better BpLabs Premium budget, and fixes from your reports. Available on Windows and Mac.

  • Voxel Plugin 2 support (new, off by default). 23 tools to create a Voxel World, stamp terrain with shapes, graphs, heightmaps, meshes and splines, edit graphs and layers, read where the ground is and place actors on it. It needs Voxel Plugin 2.0 installed for your engine version; turn the extension on in Extensions. Without Voxel installed the tools tell you what to install.
  • BpLabs Premium: the per-message budget now counts BpLabs tokens. The limit that pauses a long message used to count the model’s own tokens, which are far fewer than the BpLabs tokens Premium charges, so a 1M limit could still let a single message use 15M or more. Premium now pauses at 20M BpLabs tokens, and the pause message shows BpLabs tokens. Typing budget 40m in a Premium chat changes the limit for that chat only. A request so large that one message would cost over 10M BpLabs tokens is held back before it is sent, with a message that says what to do, and nothing is charged.
  • The Claude agent no longer shows Connected when you are signed out. The check now asks the agent itself, and shows Not signed in with the Sign in button.
  • Bug reports include the chat you are in. The report window opens on your current chat, and after a crash the dialog now has “Include the chat this happened in” ticked by default (untick it to leave the chat out). A very long chat is trimmed to its newest messages so the report always sends.
  • Niagara scratch modules can read their inputs. Custom HLSL that used a module input by name failed to compile with an undeclared identifier.
  • Landscape layers set up by the AI survive a restart. The layer info assets are now saved and assigned to the landscape properly, so painting keeps working after you reopen the project. add_landscape_layer_info can also assign a layer info you already have.
  • Building a Behavior Tree no longer crashes the editor when the tree uses Blueprint tasks, and large trees build much faster.
  • A Copy button under every AI reply, so you can copy a long answer without scrolling back up to the top.
97

V2.3.0, October 16, 2026

You choose what the AI has to ask about, and it now asks before it takes a screenshot. Available on Windows, Mac and Linux.

  • Separate switches for what the AI asks about. Settings > AI Behaviour now has Shell Command Confirmation, Source Control Confirmation, Extension Confirmation and Screenshot Confirmation next to Destructive Op Confirmation. Turn off only the ones you do not want: git and PowerShell can stay unchecked while the rest keep asking.
  • It asks before the AI takes a screenshot. A screenshot of the editor, of a play session or of a plugin panel is sent to your AI provider together with the chat, so the AI now asks first and says so. One yes covers the rest of that session. New installs start with it on; an install that already existed keeps working exactly as before and can switch it on in Settings.
  • Don’t ask again does what it says. On a shell, source control, extension or screenshot prompt it switches that setting off and saves it, so it stays off in every later session. You can turn it back on in Settings.
  • Shell and git prompts follow your settings. Before, turning Destructive Op Confirmation off still left the shell and git prompts on. If you had it off, they are off now too; if you had it on, nothing changes.
  • A bug report is no longer lost when its text has an apostrophe. “It doesn’t work” used to be dropped without any message.
  • Reporting from a running agent chat includes the conversation so far, and the report window picks your most recent chat when none is selected.
  • list_plugins always shows Ultimate Engine CoPilot and its version, and set_pie_actor_property reports the value before and after, so a change that did not land is visible.
  • Linux builds (x86_64) for UE 5.5 to 5.8. New on the Downloads page. The in-editor updater does not serve Linux yet, so download new Linux builds from the website. The MetaHuman extension is unavailable on the Linux 5.6 binary engine, and the first Linux build does not include the Don’t ask again fix yet.
98

V2.3.5, October 21, 2026

BpLabs Premium shows what it costs as it goes, attached files always reach the AI, and fixes from your reports. Available on Windows for UE 5.6 to 5.8. Mac, Linux and Windows on UE 5.5 stay on V2.3.0.

  • See what a Premium message costs while it runs. The chat shows the BpLabs tokens used so far, and how much of the message’s budget that is.
  • Summarise with BpLabs Standard. When a long Premium message pauses at its budget, one click asks BpLabs Standard to sum up where the work stands. Your chat goes back to Premium afterwards.
  • A heads-up before an idle Premium chat gets expensive. After a few minutes idle, the model no longer has the chat cached, so your next message reads the whole conversation again first. The box above the composer says roughly how many BpLabs tokens that will take before you send. It never blocks Send.
  • Switching mode no longer re-reads the whole Premium chat. Moving between Just Chat, Ask Before Edit, Auto Edit and Plan keeps the cache, and so does a chat that runs past midnight.
  • Clearer Premium limits. Running too many Premium chats or Crew workers at once now says exactly that, and the request waits and retries instead of failing straight away.
  • Files you attach reach the AI on every message, not only the first one, in every chat type including Claude Code, Codex and Copilot agents. Dropping a .md file onto the chat attaches it as text.
  • Crew workers get more room, and you can set it. A worker now gets 80 replies per checkpoint instead of 40, and a reviewer 40 instead of 30. Settings > Preferences has a Worker Round Limit and a Reviewer Round Limit for a model that needs more steps. A worker that runs out now says so on its Crew card, with a pointer to the setting, instead of asking you to type into a chat that has no message box.
  • Crew can be started from a GitHub Copilot agent chat. It used to stop with “needs a caller chat context”.
  • Compacting a chat finishes even if you switch to another tab while it runs.
  • author_gdd_interview updates the same document when you give it the same title, instead of making duplicates.
  • Crash fixes from 2.3.0 reports: importing material graphs, play-test sequences, the project scanner, selecting Blueprints in the Content Browser, deleting assets, fixing redirectors, merging actors and editor UI automation.
  • Many tool fixes from your reports. Animation Blueprints and Gameplay Cues refuse a parent class they cannot find instead of using the wrong one. MetaSound inputs that take a sound now get the sound, not its name. Reading properties shows the asset a soft reference points to even when it is not loaded. Recolouring a UI Border no longer darkens it. A play-test sequence stops cleanly if the editor stops drawing frames.
  • Web search answers fall back to web results when a direct answer is not available.