Nexus

A desktop AI companion that gives the relationship a shape.

Not "smarter chat." Not "more like a human." A specific bet: that what people want from a long-running companion is the year laid down as something they can hold — a callback that lands on the right day, a Sunday letter that names what the week actually was, a trace of how their mood and hers moved together. Local-first, single-author, slow on purpose.

English · 简体中文 · 繁體中文 · 日本語 · 한국어

Nexus is under active development. Some features are stable, others are still being polished. Feedback and issue reports are welcome.

---

Why Nexus

Most AI companions chase one of three commodities: smarter responses (foundation-model arms race), more human voice (Sesame, eleven), or more aggressive engagement (the Replika / Character.AI playbook). Nexus is none of them. It chases something the others can't easily copy because it's downstream of patience: the shape of a relationship over time, made visible.

Concretely, this means seven rituals — built on top of one another, not as a feature checklist:

1. Significance-weighted memory — not everything is remembered equally; the companion's recall is tilted toward what mattered.
2. Nightly dream cycle — overnight, conversations cluster into narrative threads and 1–3 short reflections about you. The next morning's "who you are" is updated, not reset.
3. Callback queue — a memory queued from yesterday quietly resurfaces when today's conversation has room for it.
4. Anniversary callbacks — day-30 / 100 / 365 milestones; "on this day" matches against past memory dates so a year later, that one thing you said comes back gently.
5. Morning + evening bracket — a soft check-in at the start of the day and a callback to it that night. Optional, never pushy.
6. Sunday letter — every week, she writes you a short letter naming what the week actually was. Not a digest. A letter.
7. Background errand — you can hand her a task during the day; she works on it overnight while you sleep, and delivers the result at the morning bracket.

Around these sit the boring infrastructure: 5-locale UI, 18+ LLM providers, multi-engine STT / TTS with failover, Live2D, VTube Studio bridge, MCP tools, IPC security audited end-to-end. None of that is the product. The product is what those seven rituals add up to over a year of use, and what you can hold afterwards.

This is also a single-author project, on purpose. The trade-off is real: slower than a funded team, no roadmap pressure, no quarterly OKRs. The upside is that every ritual is hand-shaped — the kind of attention you can't get from a startup that needs to grow 30% MoM. What's coming next is depth, not breadth: a long-form affect-dynamics record (your baseline mood, how it moved this month, where the two of you synchronized), grounded in published affect science (Russell's circumplex, Kuppens on emotional inertia, Welivita's empathy-intent taxonomy). Read on if that sounds right; otherwise the release notes are an honest map of what's actually shipped.

Local-first, by default

Every voice frame, every memory entry, every tool call runs on your own machine. The LLM calls themselves are the only thing that leaves your computer, and you pick the provider. You can mix and match 18+ chat providers, swap STT / TTS engines, even run fully offline with a local model + local ASR + local TTS. Nothing about Nexus assumes you trust a cloud — and there is no cloud account to make.

News

Where we are right now. v0.3.1-beta.4 is shipped (security + bug-fix). v0.4 is in active development — five rituals already merged on main (early/late bracket, Sunday letter, on-this-day callback, voice-prosody emotion loop, background errand). Next is multi-day Arc + Future-self capsule + an affect-dynamics record. v0.4.0 stable will land when the seven rituals are coherent enough to call the year-shaped artifact done. No timeline pressure — slow on purpose.

• 2026.04.26 — v0.3.1-beta.4 — audit + polish patch. Memory-store data-loss race fixed (compaction rename pattern); MCP per-tool approval (M2) + workspace:set-root approval (M3); About + Weekly Recap Settings panels; ~660 ja/ko strings translated; wake-word retry log dedup actually works now; humanized error messages. Notes.
• 2026.04.26 — v0.3.1-beta.3 — bug-fix patch. Live2D no longer broken in packaged builds (CSP unsafe-eval for pixi); thinking-mode models (DeepSeek-R1, QwQ, Hunyuan-thinking) now work in multi-turn; TTS no longer wedges voiceState on "speaking"; multimodal images preserved across turns; in-app diagnostic surface upgraded. Notes.
• 2026.04.26 — v0.3.1-beta.2 — security IPC hardening on top of beta.1. Closes chat baseUrl SSRF (H5), tightens vault enumeration limit (H4 mitigation), pins local-service probe to loopback (H8). Notes.
• 2026.04.25 — v0.3.1-beta.1. Pure installer-size fix on top of stable v0.3.0 (1.2 GB → ~250 MB by excluding unused FP32 model + git-LFS residue). Notes.
• 2026.04.25 — v0.3.0 stable released. The narrative-companion release. Significance-weighted recall + dream-cycle reflections + callback queue + anniversary milestones; relationship type declaration (friend / mentor / quiet companion); "thinking of you" OS notification; idle-motion silent gestures; smooth 2-hour day↔dusk↔night blend with bolder color contrast; Liquid Glass UI re-skin; richer weather precision. Release notes · Install.
• 2026.04.24 — v0.3.0-beta.1 / beta.2. Beta line for the relationship + memory work that landed in stable v0.3.0. Notes beta.1 · Notes beta.2.
• 2026.04.22 — v0.2.9 released. Emotional memory + 5-level relationship baseline; weather + scene overhaul; Character Card v2/v3 import; VTube Studio bridge. Tag

Older releases

• 2026.04.19 — v0.2.7. Subagent dispatcher; barge-in hardening; render-storm fixes. Notes
• 2025.04.19 — v0.2.5. Autonomy Engine V2 default-on; voice/TTS reliability pass. Notes
• 2025.04.16 — v0.2.4. Voice/TTS reliability + Anthropic prompt caching + 20+ bug fixes. Notes
• 2025.04.15 — Wake-word + VAD rewrite (Plan C): main-process Silero VAD + sherpa-onnx-node, single mic stream.
• 2025.04.14 — TTS intermittency fixes: retry / per-segment events / sender teardown.
• 2025.04.12 — Speech-interrupt architecture: echo-cancelled mic + TTS-aware dynamic threshold.
• 2025.04.10 — Hybrid memory landed: three-tier hot / warm / cold + BM25 + local vector search.
• 2025.04.01 — v0.1 opened. First playable build.

Highlights

The first three are the load-bearing structure that the seven rituals stand on. Everything below them is the practical day-to-day surface — voice, scene, integrations — that keeps the project usable while the depth accumulates.

• 💝 A relationship that has a shape, not just a length. Five levels (stranger → acquaintance → friend → close friend → intimate) with one-shot milestone instructions on each crossing. Four named sub-dimensions — trust, vulnerability, playfulness, intellectual — grow from different kinds of interaction. Pick a relationship type (friend / mentor / quiet companion / open-ended) and her tone biases without overriding her persona file. The shape is the product; the level number is just legible scaffolding.

• 🧠 Memory that dreams, not just stores. Three-tier hot / warm / cold with hybrid BM25 + vector search. A nightly dream cycle clusters the day into narrative threads, generates 1–3 short reflections about you, and queues 0–2 callbacks for next conversation. Mood-aware recall means the same memory feels different in different moods. v0.4 adds long-form affect dynamics so your baseline + monthly change + her co-regulation with you become readable artifacts, not just internal numbers.

• 🤖 An inner life with restraint. Single LLM decision call per tick, fed a layered snapshot (emotion · relationship · rhythm · desktop · recent chat) and filtered through a per-persona guardrail. She can choose to stay silent, dispatch a background research helper for tasks that would benefit from one, or run an overnight errand you handed her at lunch. The autonomy is real; it's just bounded by manual approval and budget gates so it stays trustworthy.

• 🎙️ Always-on wake word + continuous voice chat. sherpa-onnx keyword spotter beside a main-process Silero VAD over one shared mic stream. Multi-engine STT/TTS with failover, sentence-immediate streaming TTS (first audio at the first comma), 6-second first-audio watchdog, echo-cancelled self-interrupt so the pet never wakes itself up.

• 🌤️ Living scene. 14 intensity-graded weather states, continuous 24h sunlight filter, 15 AI-generated day / dusk / night scene variants. Atmospheric depth, not a static wallpaper.

• 🎭 Character Card + VTube Studio bridge. Import Character Card v2/v3 (chub.ai / characterhub compatible). Drive an external Live2D model via the VTube Studio plugin API while keeping Nexus's memory / autonomy stack.

• 🧰 Subagent dispatcher. The companion can fire a bounded research loop behind the scenes — web search or MCP tools — and weave the summary into its next reply. Capacity + daily budget enforced, with cancel + per-task progress + history.

• 🔧 Built-in tools. Web search, weather, reminders. Works with native function calling and a prompt-mode fallback for models that don't support tools.

• 🖥️ Desktop awareness. Foreground window title, clipboard, and (optionally) screen OCR. Context triggers let her react to what you're actually doing.

• 🔔 Notification bridge. Local webhook server + RSS polling — push external notifications into the companion conversation.

• 💬 Phone reachable. Discord and Telegram gateways with per-chat routing. Talk to your companion from your phone, have her respond in her own voice.

• 🌐 Multilingual UI. Simplified Chinese, Traditional Chinese, English, Japanese, Korean.

• 🔄 Provider failover. Chain multiple LLM / STT / TTS providers. When one goes down, Nexus switches without tearing the conversation down.

• 💰 Cost-aware. Built-in budget metering + Anthropic prompt caching on the system + tools prefix (30-50% input token reduction on long sessions).

• 🛠️ Diagnostics built in. JSONL log export, emotion + relationship timeline charts, 30-day cost history with per-source / per-model breakdown — all in Settings → Console.

What's new in v0.3.0

Stable. Cumulative changelog from v0.2.9: 100+ commits, ~12,000 LOC, +361 tests. Backward-compatible — pre-v0.3.0 stored state migrates transparently. Full notes: docs/RELEASE-NOTES-v0.3.0.md.

Theme	What landed
🧠 Memory does work	Significance-weighted recall, dream-cycle reflections (1–3 short observations about you), callback queue (gently surfaces a past memory in the next chat), anniversary milestones at day-30 / 100 / 365.
💝 Relationship has shape	Mood-aware recall (3 modes), 5-level milestones with first-time fire, four sub-dimensions, richer reunion framing.
🤝 Relationship has type	Pick open-ended / friend / mentor / quiet companion in onboarding or settings — biases tone without overriding SOUL.md.
💭 "Thinking of you"	OS-level notification fires after a long silence, phrased to match your relationship type. Quiet hours 23–08 hard-gated.
🎬 Alive in the corner	Idle-motion silent gestures (4th autonomy V2 action), dynamic decision cadence, first-impression curious question on early replies.
🌅 Smooth scene	2-hour day↔dusk↔night blend windows with smoothstep easing; bolder color contrast — dawn pink, golden hour deep amber, deep night cool desat.
🪟 Liquid Glass UI	Violet accent re-skin, cleaned-up toolbar, time-aware emoji greeting, panel size + position now persist across launches.
🌤️ Better weather	Hourly forecast, feels-like + humidity, day-after-tomorrow lookahead.
🧹 Polish	i18n.ts split (1842 → 588 lines), shared Settings field components, regex compile cache, async-lock dedup, build-size trim (~30–60 MB).

Earlier beta line (folded into this stable)

The beta line built the foundation that this stable polishes:

• v0.3.0-beta.1 — Three independent depth axes on the relationship system: mood-aware memory recall (VAD projection + empathy/repair/reinforce modes), one-shot level-up instructions, four sub-dimensions. Notes
• v0.3.0-beta.2 — Stability + retention pass: significance-weighted recall, dream-cycle reflections, callback queue, anniversary milestones, idle motion, dynamic cadence, Liquid Glass UI, weather precision, tray + dock icons, 7 security fixes. Notes

Axis	What it does
💝 Emotional resonance recall	The same memory feels different in different moods. Three regulatory modes (match / empathy / repair) decide whether to mirror the user's current mood or surface something to help reframe.
🎯 Relationship milestones + reunion	Crossing the 10 / 30 / 55 / 80 score thresholds fires a one-shot, understated instruction the turn you cross it. Reunions reference your last topic + how the conversation ended.
🌳 Four sub-dimensions	The flat 0–100 score now decomposes into trust / vulnerability / playfulness / intellectual, each growing from a different kind of interaction and feeding specific prompt guidance.

Full breakdown of v0.3.0-beta.1 changes

💝 Emotional resonance recall

Current mood influences which memories surface. VAD (valence / arousal) projection of the 4D emotion state, with three regulatory modes:

• Reinforce (default) — match current mood + salience
• Empathy (triggered by "陪陪我" / "listen to me" + elevated concern) — match on tone and emotional weight, surface moments where she was genuinely there
• Repair (triggered by "算了 换个话题" / "move on" or sustained severe distress) — surface distant, positive memories to help reframe

A priming ring buffer (last-3 recall centroids) prevents whiplash between unrelated moods. Intensity gating means neutral-mood turns skip the feature entirely.

🎯 Relationship milestones + richer reunion

Crossing the 10 / 30 / 55 / 80 score thresholds fires a one-shot, understated instruction the turn you cross it — telling the model to perform the shift (use a name, tease gently, be quietly vulnerable) rather than announce it. No badges, no pop-ups.

Reunions now use absence duration + last-session emotion + stored topic: a 2-day gap at friend+ gets a "weave it back in naturally" directive; a 10-day gap at close_friend+ with high prior concern prompts a gentle "did it get better?" check-in.

🌳 Four sub-dimensions under the flat score

The existing 0–100 score is now a blend of four named dimensions, each growing from a different kind of interaction:

Dimension	Grows from
Trust	Bringing problems to her, accepting her help
Vulnerability	Sharing feelings, personal history, expressed sadness (first-person only)
Playfulness	Jokes, laughter, playful teasing
Intellectual	Deep questions, debate, mutual teaching

Diminishing returns prevent runaway growth. A slow daily drift toward a low baseline means dimensions erode gently with prolonged absence. Each dimension, when notably high or low, feeds specific prompt guidance — high trust tells the companion to honor that reliance, low playfulness tells her not to force humor.

🔩 Under the hood

• Electron 36 → 41 (Node 24 ABI, macOS 12 minimum).
• 93 new tests lift the suite from 486 to 665. Five previously untested modules now covered — plugin message bus, encryption, MCP + plugin host, window manager geometry, minecraft gateway.
• Main-process audit from a five-agent review: CSP script-src tightened (unsafe-eval removed), vault file written with 0o600, child-process PID null-guard, timer-leak cleanup across mcpHost / realtimeVoice / windowManager.panelBlur.
• Shared driftToward() + classifyByPatterns() helpers in src/lib/common.ts deduplicate decay-math and regex-classification kernels across emotion + relationship code paths.
• O(n²) overlap-search in chatRuntime.trimRepeatedStreamingDelta now capped at 200 chars; useChat signature hash replaces per-turn JSON.stringify(messages); SpeechOutputSection extracts a shared TuningSlider component (-100 lines of render duplication).

---

What's new in v0.2.9

The headline of v0.2.9 was the emotional memory + 5-level relationship baseline that v0.3.0-beta.1 now extends.

• Emotional memory + relationship arc — companion remembers the feel of how you parted; 5-level progression (stranger → acquaintance → friend → close friend → intimate); per-persona memory.md files survive switches.
• Weather + scene overhaul — 14 intensity-graded weather states, continuous 24h sunlight filter, 15 AI-generated day/dusk/night scene variants.
• Character Card v2/v3 import + VTube Studio bridge — chub.ai / characterhub compatible; drive external Live2D models via the VTS plugin API.
• Pet polish — inline [expr:name] mid-sentence expression tags; 13 fine-grained mood states; per-model weighted idle fidgets; mouse-drag resize.
• Render / sync fixes — cross-window save loop, runtime-state self-feed render storm, TTS-timeout cascade, wake-word transient errors all fixed.
• Internal — Autonomy V1 deleted; CI caches sherpa models + electron-builder; sherpa bundled into Mac/Linux installers.

Full notes on the v0.2.9 release page.

Install

Pre-built installer (recommended)

Grab the latest installer from the releases page:

Platform	Asset
Windows	Nexus-Setup-<version>.exe (NSIS, unsigned — click More info → Run anyway on first launch)
macOS	.dmg or .zip (unsigned — see macOS steps below)
Linux	.AppImage / .deb / .tar.gz (AppImage: chmod +x then run)

macOS first launch (unsigned build)

1. Open the .dmg and drag Nexus.app into /Applications.
2. Remove Gatekeeper's quarantine flag — open Terminal and run:

   xattr -dr com.apple.quarantine /Applications/Nexus.app

   (Or: right-click Nexus.app → Open → confirm in the dialog.)
3. Launch Nexus. On first run a "安装本地语音模型" wizard appears — click 一键下载 to pull ~280 MB of sherpa-onnx + VAD models into ~/Library/Application Support/Nexus/sherpa-models. The wizard can be dismissed and reopened later from Settings.
4. Python-based options (OmniVoice TTS / GLM-ASR) are detected automatically. If you haven't installed Python + requirements.txt, they're silently skipped — the core chat + SenseVoice STT + Edge TTS stack still works.

After the wizard, a 4-step onboarding guide walks you through: persona, main chat model, voice stack, companion preferences. You can skip any step and adjust in settings later.

Build from source

Requirements: Node.js 22+, npm 10+. (macOS: Xcode Command Line Tools for native modules.)

git clone https://github.com/FanyinLiu/Nexus.git
cd Nexus

# Windows:
setup.bat

# macOS / Linux:
bash scripts/setup.sh

# Dev mode with hot reload
npm run electron:dev

# Production installers
npm run package:win      # → release/Nexus-Setup-<version>.exe
npm run package:mac      # → release/Nexus-<version>.dmg (universal build via electron-builder --arm64 --x64)
npm run package:linux    # → release/Nexus-<version>.AppImage / .deb

Production installers end up in release/. Signing is off by default (unsigned macOS builds require right-click → Open on first launch to bypass Gatekeeper). Wire in your Apple Developer ID / Windows signing cert via the usual electron-builder env vars (CSC_LINK, CSC_KEY_PASSWORD, APPLE_ID, APPLE_APP_SPECIFIC_PASSWORD, APPLE_TEAM_ID).

macOS permissions. On first launch, Nexus will prompt for:

• Microphone — required for voice conversation / wake word / STT.
• Screen Recording — required for desktop context / OCR. Approve in System Settings → Privacy & Security → Screen Recording, then restart Nexus.
• Automation — optional; used by Now Playing (Music / Spotify) and foreground-app detection. If denied, the relevant features silently fall back to empty state.

macOS packaging notes. The mac .dmg ships without bundled sherpa models (Windows / Linux installers still bundle them). The in-app setup wizard downloads them to ~/Library/Application Support/Nexus/sherpa-models on first launch. This keeps the .dmg ~250 MB instead of ~550 MB and survives app upgrades (downloaded models persist across updates since they live in userData, not inside the .app bundle).

Configure

After first launch, open Settings:

• Chat — pick a provider, paste your API key, choose a model. Supports chained failover (primary + N fallbacks).
• Voice input — choose STT engine (local SenseVoice or sherpa runs fully offline; cloud options include Zhipu GLM-ASR, Volcengine, OpenAI Whisper, ElevenLabs, Tencent). Set wake word + VAD sensitivity here.
• Voice output — pick a TTS (Edge TTS is free and fast; MiniMax / Volcengine / DashScope for natural voices; OmniVoice for on-device). Streaming enabled by default.
• Memory — dream cycle cadence, recall depth, embedding model.
• Autonomy — emotion / relationship / rhythm tuning, proactive greeting thresholds.
• Integrations — Telegram / Discord bot tokens, notification webhook port.

Supported providers

Category	Providers
Chat (18+)	OpenAI · Anthropic · Gemini · DeepSeek · Kimi · Qwen · GLM · Grok · MiniMax · SiliconFlow · OpenRouter · Together · Mistral · Qianfan · Z.ai · BytePlus · NVIDIA · Venice · Ollama · Custom
STT	GLM-ASR-Nano · Paraformer · SenseVoice · Zhipu GLM-ASR · Volcengine · OpenAI Whisper · ElevenLabs Scribe · Tencent ASR · Custom
TTS	Edge TTS · MiniMax · Volcengine · DashScope Qwen3-TTS · OmniVoice · OpenAI TTS · ElevenLabs · Custom
Web search	DuckDuckGo · Bing · Brave · Tavily · Exa · Firecrawl · Gemini Grounding · Perplexity

Recommended model setup

These recommendations target English-speaking users. For other languages see 简体中文 · 繁體中文 · 日本語 · 한국어.

Chat model (LLM)

Use case	Provider	Model	Notes
Daily companion (top pick)	Anthropic	claude-sonnet-4-6	Best overall quality, stable tool calling, natural English
Daily companion (budget)	DeepSeek	deepseek-chat	Extremely cheap, good multilingual, great for long conversations
Budget friendly	OpenAI	gpt-5.4-mini	Fast and cheap, solid English, good for high-frequency chat
Free tier	Google Gemini	gemini-2.5-flash	Generous free quota, good for getting started
Deep reasoning	DeepSeek	deepseek-reasoner	For complex reasoning, math, and code

Speech-to-text (STT)

Use case	Provider	Model	Notes
Best accuracy	OpenAI	whisper-large-v3	Industry standard, highest English recognition accuracy
Budget friendly	OpenAI	gpt-4o-mini-transcribe	Multilingual, works with existing OpenAI key
High-accuracy cloud	ElevenLabs Scribe	scribe_v1	99 languages, excellent punctuation and speaker detection
Local streaming	Paraformer	paraformer-trilingual	Real-time transcription while speaking, low latency
Local fast	SenseVoice	sensevoice-zh-en	15× faster than Whisper, offline

Text-to-speech (TTS)

Use case	Provider	Voice	Notes
Free (recommended)	Edge TTS	Jenny (en-US-JennyNeural)	Microsoft free, warm American English female voice, no API key
Free (male)	Edge TTS	Guy (en-US-GuyNeural)	Calm American English male voice, free
Best quality	ElevenLabs	Custom voice_id	World-class speech synthesis, voice cloning supported
Cloud general	OpenAI TTS	nova / alloy	Works with existing OpenAI key, gpt-4o-mini-tts model
Local offline	OmniVoice	Built-in voices	Fully offline, local port 8000, runs on RTX 3060

Architecture

Layer	Technology
Runtime	Electron 41
Renderer	React 19 · TypeScript 5.9 · Vite 8
Character	PixiJS 6 · pixi-live2d-display
Voice (client)	WebAudio · sherpa-onnx-node · Silero VAD · Web Speech API fallback
Voice (server)	Local OmniVoice / GLM-ASR-Nano sidecars over HTTP
Local ML	onnxruntime · @huggingface/transformers
Storage	localStorage · vault-encrypted API keys · SQLite-style JSON memory store
Packaging	electron-builder · electron-updater

Higher-level layout:

src/
├── app/            # Top-level views, controllers, overlays
├── features/       # Voice, chat, autonomy, tools, memory, agent
├── hooks/          # React hooks (voice / chat / reminders)
├── components/     # Reusable UI
├── lib/            # Storage, runtime bridges, plain helpers
└── i18n/           # Locale bundles
electron/
├── main.js         # Entry
├── ipc/            # Typed IPC handlers
├── services/       # TTS / STT / tools / key-vault
└── sherpa*.js      # On-device voice engines

Roadmap

Planned

• Screen-aware proactive conversation — periodically read screen context (foreground app, visible text) and initiate conversation about what the user is doing, not just respond when spoken to.
• Decision / Roleplay / Agent three-layer separation — split intent classification (fast) from roleplay (persona-pure) from background agent tasks. Roleplay never sees tool metadata; agent results are "announced" by the character in its own voice.
• Character diary & autonomous timeline — the companion auto-generates a first-person diary entry each day summarizing what happened; optionally posts "moments" to a browsable feed, creating a sense of independent life.
• Daily schedule & activity states — the companion follows routines (work / eat / sleep / commute) that affect availability, tone, and energy. Late-night conversations feel different from morning ones.
• Mini mode / dock-edge hide — drag the pet to the screen edge and it auto-hides with a peek-on-hover animation. "Always present, never intrusive."
• Webcam awareness — use MediaPipe face mesh to detect fatigue signals (yawning, eye closure, frowning) and inject detected state into the companion's context so it can proactively react.

Ongoing

• Pipecat-style frame pipeline replacing the monolithic streaming TTS controller (Phase 2-6; Phase 1 shipped in v0.2.4).
• Auto-update infrastructure via electron-updater + signed binaries.
• Mobile companion app (voice-only remote for the desktop instance).

Community

Nexus is a solo-maintained project, which means issues and PRs move faster when the triage channel matches the question:

• 🐛 Found a bug? → Bug Report
• 💡 Small, well-scoped feature idea? → Feature Request
• 🧠 Bigger or open-ended idea? → Ideas Discussion first, so others can weigh in before it becomes a tracked task
• ❓ Stuck on setup or usage? → Q&A
• 🎨 Want to show how you use Nexus? → Show and tell
• 💬 Just want to chat? → General
• 📣 Release notes and roadmap updates → Announcements

Contributing

Contributions welcome — bug fixes, new providers, UI tweaks, translations, Live2D models, or new autonomous behaviors. Even a one-sentence issue or a typo-fix PR moves things forward.

Quick start:

• Read the full Contributing Guide for development setup, project layout, code style, and PR workflow.
• Use the issue templates for bugs and feature requests — they keep reports consistent enough to triage quickly.
• Run npm run verify:release (lint + tests + build) before pushing — this is exactly what CI runs.
• Follow Conventional Commits for messages: feat:, fix:, docs:, refactor:, etc.
• One logical concern per PR. Split unrelated fixes into separate PRs.

All participation is governed by the Code of Conduct — short version: be kind, assume good faith, focus on the work.

Security issues

If you find a security vulnerability, please do not open a public issue. Open a private security advisory instead.

Star history

License

Released under the MIT License.