Mana — local-first open-source AI companion for Windows (Yuuzulight, 2026)
Published · Researched 2026-09-20
Status: Live, open-source, under active development as of September 20, 2026 (repo README references a native-launcher parity milestone dated 2026-09-21). Apache-2.0. Not a commercial product and not crowdfunded — included as a genuine 2026 local-first 3D-companion software launch. First-release date UNVERIFIED. Research current as of September 20, 2026.
Overview
Mana is a local-first AI companion for Windows by developer Yuuzulight: local speech transcription (Whisper), local LLM replies (Qwen3/Qwen2.5 GGUF profiles via llama.cpp), and local TTS (Fish Speech, Kokoro, GPT-SoVITS), with screen awareness, memory, web access via a local SearXNG instance, and editor integration (Zed/VS Code handoff, Agent Client Protocol). The avatar layer: a built-in Live2D VTuber avatar with lip sync, PNG avatar overlay, VTube Studio hotkeys/reactions, optional VRM 3D avatar support in the legacy launcher, and a vision model option (Qwen2.5-VL) for image understanding. A native C#/WinForms launcher reached full feature parity with the legacy Electron launcher in September 2026 and is now the primary launcher; a mobile PWA via Cloudflare Tunnel, Discord bot, voice barge-in, and MCP support are on the roadmap. Remote AI is explicitly opt-in (MANA_ALLOW_REMOTE_AI=1); default posture is fully local.
Social standing
Developer-community only. No Facebook groups, Threads posts, Instagram content, or end-user reviews were found in this pass — the social layer is the GitHub repo itself (issues, discussions, docs). The README documents an active, meticulous solo-dev workflow: a 5-round independent verification pass on the native launcher, measured memory benchmarks, per-feature setup docs, and a "deep research" roadmap. This is a hacker's companion project, not a consumer product with a fandom.
Specs
| Area | Detail |
|---|---|
| 3D tech | Live2D avatar with lip sync (built-in); PNG avatar overlay; VTube Studio integration; optional VRM 3D support in legacy launcher |
| Voice | Local STT (Whisper) + local TTS (Fish Speech default, Kokoro, GPT-SoVITS); voice barge-in on roadmap |
| Text chat | Yes — OpenAI-compatible local API |
| Memory | Yes — memory system plus local data storage docs |
| Screen awareness | Yes — local vision model describes screen on hotkey (Ctrl+Alt+M); OCR endpoint |
| LLM | Local GGUF profiles: Qwen3-4B default, Qwen3-14B/8B quality, Qwen2.5-coder-7B coding, Qwen2.5-1.5B fallback |
| Privacy | Local-first; remote AI requires explicit opt-in; local web search (SearXNG) |
| Platforms | Windows (native C#/WinForms launcher primary; legacy Electron fallback); mobile PWA via Cloudflare Tunnel (roadmap/active) |
| License | Apache-2.0 |
Pricing (as reported, dated)
Observed September 20, 2026:
- Free — open-source, Apache-2.0, no price, no subscription, no crowdfunding. Costs are the user's own hardware/electricity and optional remote-AI API keys.
- No monetization found — UNVERIFIED whether any paid tier is planned.
Support reality
- Support = GitHub Issues and GitHub Discussions, per the README;
npm run doctordiagnostics built in. - Solo-developer project: bus-factor risk is the support reality. No SLA, no commercial support channel.
- Actively maintained through the observation window (September 2026 milestones documented in-repo).
Community verdict
The most technically serious local-first companion in this round — and the only one where "your data stays home" is an architecture rather than a slogan. But it is developer software: Windows-only, self-hosted models, README-driven setup. As a 2026 3D-companion launch it qualifies on activity and recency, not on consumer polish. No end-user community exists to quote.
Primary images
mana-01.jpg— official README screenshot: Mana windows launcher UI (converted from PNG to JPEG).mana-02.jpg— official README screenshot: Mana desktop client main UI (converted from PNG to JPEG).
Both verified JPEG (file-type check, ffmpeg conversion), content-accurate — official project screenshots, no substitutions. Only two images: the repo README contains just these two UI screenshots (plus an architecture SVG); no avatar-render screenshots were available — documented, not dropped.
Evidence status
- Local-first architecture, model stack, Live2D/VRM avatar support, launchers, roadmap, Apache-2.0, September 2026 activity — GitHub README (Yuuzulight/Mana), verified.
- AlternativeTo lists it as a Replika alternative (12th, Freemium tag is AlternativeTo's label; the project itself is open-source) — verified listing.
- UNVERIFIED in this pass: first-release date, contributor count, download/user numbers, any end-user reviews, Facebook/Threads/Instagram/Reddit community, paid-tier plans.