Create, stream, remember the room, and grow while you sleep. One local-first control room for Live2D talent that can make media, run a real show, and learn what worked, nya~
28 feature families spanning creation, live intelligence, community, growth, provenance, and operations â self-hosted and yours.
Headless browser-based Live2D recording â no GPU needed. Trigger expressions with emotion markers (*happy*, *surprised*), add TTS voice, background music, and auto-burn subtitles into the final MP4.
Upload a song â separate vocals with Demucs â convert voice via RVC API â mix with FFmpeg â generate lip-sync data â render the final video. Karaoke creation and LRC/MusicXML notation supported.
Put 2â4 Live2D talents on one stage. A floor director balances airtime, obeys direct mentions, keeps pair threads alive, and caps consecutive turns â alongside real-time Twitch & Kick chat, speech queuing, and an OBS browser source. NEW: live follow goals with fireworks, sprite girls greeting every follower, and a daily auto-stream window so the show runs while you're literally asleep uwu.
Rebuilt in August 2026: OpenCV motion and HUD deltas decide when something happened, a vision model reports only what it can verify, and plain Python decides whether that earns a line. She lands it on the exact second and is not allowed to name a weapon the HUD never showed.
Generate character art via Stable Diffusion with full control: sampler, width/height, CFG scale, seed, batch sizes (1, 4, 10, 25, 50, ð¯), HR upscaling, face restore, meme mode with top/bottom text, and auto emotion detection.
Auto-find the best frame with AI, apply styles (default / vibrant / dramatic / soft / neon), add text overlays, compare styles side-by-side, and batch generate for your entire queue at once.
Track YouTube Videos, Instagram Posts, Total Views, Total Likes, Total Comments, and Engagement Rate. See top 3 performing content by platform, monthly activity, best posting times, and top emojis.
Real-time SEO score (A/B/C/D grade), keywords found, hooks found, emojis found. AI generates viral titles with style presets, analyzes descriptions (word count, CTAs, timestamps, hashtags), and suggests keywords.
Filter by region (US, Japan, UK, KoreaâŠ) and niche (VTuber, Gaming, Anime, React, ASMR, Music). See rising queries, virality scores (High/Medium/Low), and one-click create content from any trending keyword.
Unified inbox for unanswered comments & mentions across YouTube, Instagram, and Twitter. AI drafts replies, tracks follower deltas per platform, and lets you mark interactions answered with one click.
Log donations from YouTube, Twitch, and Fanbox. USD-normalized revenue comparison, monthly growth charts, currency mix, top supporter rankings, 3-month forecast with trend indicator, and CSV/JSON exports.
Create users, assign roles (Creator â Editor â Manager â Moderator â Admin), reset passwords, and configure CORS allowed origins. Rate-limited logins with Discord alerts after 3 failed attempts.
At 2 AM JST, Daily mode cooks tomorrowâs three Shorts plus three social posts for every connected platform. Buffer mode keeps a ready-to-publish stash so the calendar never looks abandoned.
Transcribe long footage, correct captions, generate chapters, detect spicy highlight moments, and turn a stream-sized monster into short-form clips with platform-ready packaging.
Send heavy renders to batch workers with idempotency keys, renewable leases, heartbeats, and abandoned-job recovery. Watch scheduler health and metrics, retry failed uploads with backoff, and ping Discord when a pipeline starts acting possessed.
A persistent showrunner moves each stream from cold open to payoff and callback, watches audience energy, scores spoken beats, and starts a fresh mini-episode instead of drifting into endless AI small talk.
Recognize returning viewers with semantic memory and relationship dimensions while keeping identity linking consent-gated. Memories carry provenance, confidence, and temporal validity, and newer facts supersede stale ones. Moderators can inspect or erase memories and bond levels independently.
Drive one Live2D model from webcam blendshapes while an AI co-host listens, answers, emotes, and yields when interrupted. Real visemes, turn cancellation, provider-swappable voice and full WebM recording included.
Shared motion timing aligns visemes, prosody, emotions, blinks, gaze, breathing, and physics. Listening characters add tiny cached âmhmâ and âoh!â backchannels without stealing the speakerâs turn.
Platform-aware daily caps, idempotent scheduling, honest automated-review trails, synthetic-media disclosure, failure visibility, and pipeline smoke tests keep overnight generation ambitious without becoming reckless.
At 23:30 JST every talent writes a first-person entry built only from the day's real database events â what published, what you approved or rejected, what failed to upload, who donated, which streams ran. Entries become memories, so a day she actually had can resurface on stream weeks later.
The AI disclosure is drawn into the scene itself, not just the page, so it survives the capture canvas, a direct RTMP push, and an OBS Virtual Camera. Final renders can be C2PA-signed with c2patool, and every generated asset carries an auditable recipe with asset hashes â no raw prompts, no secrets.
Every AI-generated post is compared against 60 days of the same talent's own output in the same format before it can leave the queue. Cross 0.92 similarity and it is held, not published â because the inauthentic-content policies do not care that a template was technically new.
!remember files an explicit, high-importance fact with the whole cast present. !correct retires the wrong version instead of deleting it. !memories reports what is on file. !forget is admin-only, arms rather than fires, and waits for confirmation.
Put a talent into random video chat as an actual webcam and an actual microphone: OBS Virtual Camera carries her face, one loopback device carries her voice in, a second carries the stranger's voice out to Whisper. Preflight refuses to call the bridge ready when her voice and her ears share a device.
One button checks recording disk, FFmpeg, RTMP destinations, Twitch EventSub, Kick webhook readiness, OBS connectivity, active FPS, skipped frames, output congestion, and simulcast overlay policy. Optional services fail visibly without taking the core stream hostage.
A domain layer that imports nothing from FastAPI, Live2D, or OBS holds canonical identity, personality, memory authority, and typed events. Dialogue emits a PerformanceIntent; Live2D, VRChat OSC, and Discord adapters consume the same port. She stays herself wherever she is embodied.
Every attempt is recorded with a verdict on whether the returned URL identifies a post or merely an account â the difference that let a six-month Instagram outage hide behind 411 posts marked success. Per-account rate limits and durable job leases come with it.
A supervised pipeline â from blank canvas to disclosed, review-ready content for every platform.
The entertainment brain gives every livestream an episode arc, persistent character state, audience awareness, and a memory of what made the room light up. The result is a show that accumulates meaning instead of an endless bucket of unrelated one-liners.
/metrics.Every night at 23:30 JST each talent writes a first-person entry built only from the day's real database events. What published. What you approved or rejected. What failed to upload. Who donated. Which streams ran. How many videos she was made to record.
/diaryThe old âgive a model a bag of frames and ask for eight beatsâ pipeline was deleted in August 2026. It was fluent and it was making things up. The Reactive Engine replaced it with five separated stages, and no model is called until the pixels say something happened.
Real-time SEO scoring, top-performing content by platform, best posting times, and AI title suggestions â all before you hit publish.
2â4 talent streams now have a floor director, episode direction, audience-aware interrupts, natural listening reactions, resilient multistream output, and a daily auto-stream window. We are SO automated it's almost rude, nya~ ð¶
!remember, !correct and read !memories â but taking over the show needs an admin dashboard account matched by platform handle, per platformMonitor what's rising right now by region and niche, then generate on-trend art and thumbnails instantly.
Daily mode makes tomorrow happen. Buffer mode hoards Shorts for later. Both modes are supervised by a tiny, very determined scheduling goblin.
Three vertical videos, then tailored posts for every platform you actually configured â held in the editorial queue until a human gives the little monsters a final nod.
The cute buttons are backed by job tracking, remote batch workers, retries, health checks, metrics, and alerts. You can make a media empire without staring at a spinning tab like it owes you rent.
/health, /metrics, and structured logsThe EU AI Act's transparency obligations went live on 2 August 2026, and the platform policies now name mass-produced templated output directly. So the disclosure is drawn into the scene, provenance travels with the file, and a similarity gate stops the overnight renderer from quietly plagiarising yesterday.
c2patool, signed after the last FFmpeg pass so the manifest survives subtitle burn-in and the audio muxcontainsSyntheticMedia is forwarded on YouTube upload and the disclosure text reaches the description instead of dying in a database columnHook up OBS, Stable Diffusion, Discord webhooks, PostgreSQL, and email reports â all from the Settings page.
Create and manage multiple VTuber profiles, each with their own platform credentials, Twitch bot, Discord persona, and per-VTuber analytics.
5 permission levels â Creator â Editor â Manager â Moderator â Admin â so your whole team can collaborate safely.
A carefully chosen stack that prioritizes async performance, AI extensibility, and zero cloud lock-in.
Self-hosted, wildly extensible, and held together by taste, telemetry, and a mildly concerning amount of FFmpeg. Your content. Your rules. Your tiny empire.