Steam Community :: AI Desktop Pet

AI Desktop Pet

All Discussions Screenshots Artwork Broadcasts Videos Workshop News Guides Reviews

AI Desktop Pet

8 in Group Chat

An all-in-one Live2D desktop pet, local tavern-style AI roleplay and visual-novel experience. Built-in local LLM runs without API keys or internet, with voice chat and voice cloning, broad compatibility with mainstream persona cards, and Steam Workshop support.

$12.99

Visit the Store Page

Most popular community and official content for the past week. (?)

v1.1.022 — 2026-05-16

12:13am

## v1.1.022 — 2026-05-16
- feature: TTS "Read aloud" mode — new dropdown in the Voice (TTS) settings panel that controls how much of the AI's reply gets spoken. **All text** reads the full response (existing behavior). **Dialogue only** reads just the contents of quoted dialogue and skips surrounding narration / scene description, which is especially useful for roleplay characters whose replies mix prose around the spoken line. The setting takes effect on the next sentence and persists per profile.

- feature: Pet speech-bubble display mode — new tri-state setting in the Voice (TTS) panel that decouples what the pet says aloud from what its on-screen bubble shows. **Full text** shows the complete AI reply (the default, matching prior behavior). **Spoken only** shows just the words currently being read aloud — pairs naturally with "Dialogue only" when you want only the actually-spoken text on screen. **Hidden** turns the bubble off entirely. The previous "show chat bubble" on/off toggle migrates automatically: off becomes **Hidden**, on becomes **Full text**.

- feature: "Drop action blocks" toggle in the Voice (TTS) panel. When enabled, asterisk-wrapped action descriptions like \`*smiles softly*\` are removed entirely from the TTS pipeline rather than just stripping the asterisks and reading the surrounding words. Off by default.

- improvement: with **Full text** + **Dialogue only** in danmaku-bubble mode, narration segments now also float up as their own bubble alongside the spoken-dialogue bubbles, instead of going unseen. Each narration bubble's on-screen lifetime scales to text length and adapts to script — character-based scripts (Chinese, Japanese, Korean), no-space Southeast Asian scripts (Thai, Lao, Khmer, Myanmar), and space-delimited scripts (Latin, Cyrillic, Greek, Arabic, Hebrew, Devanagari, Vietnamese) each use a rate appropriate to how that script reads. Bubble lifetime is bounded between 2.5 and 12 seconds so short bubbles still register and very long passages don't pile up.

- improvement: pet speech bubble reliability — the bubble now clears cleanly at the start of every new AI turn, danmaku-mode floating bubbles display each segment independently so they no longer overlap or visually merge, and the dialogue extractor is more tolerant of AI replies that mix curly ("...") and ASCII ("...") quotation marks across the same response, so the spoken text matches what's between the quotes even when the model is inconsistent.

- fix: AI chat could fail to initialize on AMD and Intel GPUs in the packaged Steam build, with the chat status surfacing "No GPU backend available. LLM requires GPU (Vulkan or CUDA)" even on hardware where Vulkan was fully supported by the driver. The Vulkan inference path was looking for its runtime binaries at a location inside the application archive that Windows cannot load native modules from — the same packaging issue that was previously fixed for the CUDA paths in v1.1.018, now also resolved for Vulkan. AI chat now reliably comes up on Radeon RX 5000-series and newer, Intel Arc, and any other Vulkan-capable GPU on the first launch. NVIDIA users were unaffected by this issue and remain unaffected.

- improvement: the GPU backend cascade (CUDA 13 → CUDA 12 → Vulkan) is more resilient to driver- and engine-level error messages. Previously, a small class of transient error texts emitted by certain GPU drivers could cause the cascade to stop early and report failure without trying the remaining backends. The cascade now correctly proceeds to the next backend in these cases, giving Vulkan a chance to take over from a failing CUDA attempt on mixed-vendor or oddly-configured systems.

- improvement: expanded the list of older NVIDIA GPUs that are recognized as pre-Volta up front, so they skip the CUDA 13 attempt and go straight to CUDA 12 on first launch. Added coverage for GeForce GT 1030 (Pascal), GT 740 / 730 / 720 / 710 (Kepler / Fermi), GT 640 / 630 (Fermi), GTX 690 (Kepler dual-GPU), and Titan Black / Titan Z (Kepler). Users on these cards no longer experience the brief first-launch crash before the runtime blacklist (added in v1.1.019) catches the mismatch — CUDA 12 is selected directly.

AI Desktop Pet

模型问题

将渲染质量调高会导致人物不见，调整位置也没有用，只能露出一小部分

Rem0221

3 hours ago

General Discussions

v1.1.021 — 2026-05-15

May 15

## v1.1.021 — 2026-05-15
- feature: edit any chat message after the fact — both your own and the AI's. Every chat bubble now has an Edit button in the hover action bar (alongside Copy, and for AI messages alongside Regenerate as well). Edit opens an inline text area where you can Cancel, **Save** the edit (rewrites the stored message in place without re-asking the AI — same effect as the right-click → Edit action in the Memory Manager window, just reachable directly from the conversation), or for your own messages, **Send** to truncate the conversation at that point and re-ask with the new text. Edits persist across restarts via the chat history store, and work on historical messages from previous sessions, not just the most recent turn.

- feature: bulk-delete all chat records for the active character in one click — new "Delete all" button placed immediately to the right of "Clear today" in the Memory Manager → Chat panel. Asks for confirmation (with text emphasizing the action cannot be undone) before removing every chat record for that character across all dates. Persistent memories — the extracted knowledge entries on the right side of Memory Manager — are not affected and remain managed separately by the "Clear auto memories" button as before.

- feature: VN-mode dialogue box now shows the same per-message action bar as the chat window — Copy, Edit, and Regenerate (the last one on the AI's reply). Buttons reveal on hover. Edit/Save/Send/Regenerate all delegate to the same backend handlers as the chat window, so anything edited or regenerated in VN mode stays in sync with what the chat window sees and what gets saved to the chat history store. The full-history LOG overlay shows only the Copy button per entry, because the LOG is a read-only past-conversation snapshot and adding mutating actions there would create a desync between what the LOG shows and what the live conversation actually contains — Copy is the only useful action in that context.

- improvement: the inline message edit area is now much roomier. Previously the edit text box inherited the chat bubble's narrow column (~85% of the chat list width in normal mode, capped at 720 px in maximized mode), and the textarea itself opened at only ~40 px tall (about one line of text), which made editing a multi-paragraph AI reply cramped. The edit text area now expands the message row to roughly the full chat-list width (up to 1100 px in maximized mode), and the textarea starts at 120 px tall (~6 lines) and can grow up to 400 px. The same change applies to VN-mode dialogue-box edits.

AI Desktop Pet

v1.1.020 — 2026-05-15

May 15

## v1.1.020 — 2026-05-15
- feature: Spine model support extended to cover Spine 1.6 through 4.3. Previous releases supported Spine 3.7 through 4.2 only; older models from the 2014–2018 spine-js era (used heavily in legacy game rips, older VTuber bundles, and a large amount of community content from that period) and the very recent 4.3 beta exports now load and animate correctly. The application picks the matching rendering backend automatically based on each model's embedded version — 2D canvas for Spine 1.6–3.3, standalone WebGL for 3.4–3.6 and 4.3, pixi-spine for 3.7–4.2 — with no user-visible difference between the three paths beyond compatibility. Models whose version is detectable but falls outside the supported range now show a clear notification stating the detected version and what is supported, instead of silently falling back to the default character.

- improvement: built-in VRM library curated to a tight set of six character models — Avatar Anne, Victoria Rubin, Sendagaya Shibu, Sendagaya Shino, Vivi, and Vita. Display names in the model picker now show just the character name; the full license information for each model continues to surface in the detail panel and in the third-party notices. The pre-cleanup library shipped eleven models, several of which were redundant alternates of the same character or VRM SDK feature-demo models with no real artwork. Reduces installation footprint by approximately 70 MB.

- improvement: built-in Shimeji set simplified to the Slime character. Third-party Shimeji directories remain fully importable through the import button — the change only affects which characters come pre-installed.

- improvement: clicking a different model-type sub-tab in the Settings pet picker (Live2D / VRM / Spine / Shimeji / Static) now switches the running pet to that mode, picking up the most recently used model of that type. Previously the sub-tab clicks only changed the browse view inside Settings without affecting the running pet, so users had to additionally click the model card to make the switch take effect.

- fix: importing a model of any type (Live2D / VRM / Spine / Shimeji / Static) now consistently switches the main pet window to the newly imported model and updates the Settings detail panel to point at it on the first try. No need to click the new card or restart afterwards for the import to "take".

- fix: newly imported Live2D models no longer briefly appear stretched or squished before settling into shape on first display. The import path now follows the same clean rebuild used by manual model switching, so the canvas dimensions and the model's aspect ratio are in sync from the first rendered frame.

- fix: Spine pet's feet and lower body no longer get cut off at the bottom edge of the window. Tall Spine models — including standard idle animations that step feet slightly downward, as found in most Arknights-style character models and in the bundled spineboy — now display fully within the visible area.

- fix: opening the Settings window no longer briefly re-renders the currently displayed pet. The momentary refresh was most noticeable on Spine pets, which would visibly flicker on Settings open; Settings now opens with the running pet untouched.

- fix: switching between Spine models with substantially different aspect ratios — for example, the tall narrow spineboy to the square alien demo — no longer visibly shifts the pet's screen position to one side. Programmatic window size changes now preserve the visual center of the pet rather than anchoring at the top-left corner of the window, so the on-screen position stays put across both model swaps and across moves of the size slider.

- fix: closing the application and re-opening it now restores the pet to the exact on-screen position it was at when last closed. Earlier behavior could place the pet some distance up and to the left of its last visible location on each launch.

AI Desktop Pet

[Official / Pinned] v1.1.0 — Bug Reports, Suggestions & Feature Requests

Hi everyone, I'm the solo developer behind this game. Thank you for trying v1.1.0! To make future updates better, I've opened this thread as the central place for all feedback - bug reports, improvement ideas, and feature requests are all welcome. I read every reply personally. Please use the following format when posting: Bug / Suggestion / Feature request Windows version / GPU / VRAM [App version] v1.1.0 (Settings >> About) [Steps to reproduce] (for bugs - the more detail, the better) ...

伤风爆

3 hours ago

General Discussions

API Key plz.

I noticed that there currently isn't an option to input a custom API key. I would like to kindly request that you consider adding a feature that allows users to register their own API keys (such as Gemini Pro(Ai Studio), OpenAI(Codex), etc.). If users can input their own API keys, we would be able to use more advanced AI models and manage our own API usage/costs. I believe this feature would make your app even more incredible and flexible for many users. Thank you for your hard work and time. I look forw...

2 hours ago

General Discussions

Workshop Item

【Live2D/VTS】Nahida_1080（ナヒータ、纳西妲）

伤风爆

View Steam Workshop items

Workshop Item

Leila

luparia

View Steam Workshop items

Workshop Item

Zero Two

Devorain

View Steam Workshop items

Workshop Item

Hikari & Nozomi

luparia

View Steam Workshop items

Workshop Item

Furina

luparia

View Steam Workshop items

API

什么时候会更新接入API呢，本地又跑语音又跑模型的，有点带不动了

Rem0221

May 15 @ 4:41am

General Discussions

What is this?

https://imgur.com/a/why-YEmLmOV The pet is stuck on Loading conversation model - Intiali... 20% No GPU backend available. LLM requires GPU (Vulkan or CUDA). (CUDA: NoBinaryFoundError) Is my GPU bad? I got a RX5700XT and RYZEN 5 3600x

Ledroy

12 hours ago

General Discussions

See More Content

No more content. So sad.

You can help: share a screenshot, make a video, or start a new discussion!

Returning to where you were last on this page...

Cancel