Selene

Selene is an agent-first desktop app that runs AI on your machine. Chat, write code, generate images, design UIs, control a browser — then pipe all of it into WhatsApp, Telegram, Slack, or Discord. Your data stays on your device. Every part of Selene — chat, embeddings, voice, images — lets you pick between local and cloud. Run fully offline or bring your own API keys. Mix and match.

Agent-First, Not Button-First

Every UI action is also an agent action — step in anywhere or let it run end-to-end.

Why We Built It

AI is expensive because models re-read everything, every turn. Selene runs a small retrieval pipeline first — it finds what's relevant, the main agent picks it up and moves on. Tools load on demand. Context stays lean. You pay less per turn.

What's New in v0.3.4

• Voice input, rewritten. Your words land in the composer the instant recording stops; grammar polishing happens in the background. Start a new recording while a previous one is still polishing — they no longer block each other.
• Design Workspace. Generate UI components with AI, preview them live in an isolated sandbox, edit-in-place via patches.
• Ghost OS. Agents can see your screen through a vision sidecar, wired into the MCP tool pipeline.
• Folder Sync. Replaces the old Knowledge Base. Native FSEvents on macOS, toast notifications, and a proper progress bar.
• Claude Opus 4.7 + Kimi OAuth. New flagship Anthropic model (1M context, thinking) and device-flow sign-in for Kimi — no more API-key copy-paste.
• Platform bump. Electron 41 / Chrome 146 / Node 24, 9 Dependabot advisories closed, hundreds of unused files pruned.

Full notes: RELEASE_NOTES_v0.3.4.md · All releases on GitHub

Modes

Selene Dev

• Git, diffs, and PRs. Stage, branch, diff, PR — from the UI or via agent.
• Built-in browser. Agent-controlled Chromium with console log access and session replay.
• Output protection. Bundled Rust tool trims long build/test output before it hits the model.
• Automatic checks. Hooks run type-checking, linting, or any custom logic after agent edits.

Selene Fun

• 3D avatar. Animated face with lip-sync and emotion detection.
• Voice cloning. Custom voice via ElevenLabs or Microsoft.
• Scheduled assistants. Cron tasks delivered to any connected channel.
• Memory. Selene surfaces things to remember; you approve what sticks.

Design Workspace

• AI components, live. Generate React/HTML components; they render immediately in an isolated sandbox.
• Edit-in-place. Ask for tweaks and the agent applies targeted patches instead of regenerating from scratch.
• Responsive previews. Mobile / tablet / desktop viewport toggles, plus Light / Dark / System.

Chromium Workspace

• Real browser, agent-driven. Embedded Chromium controlled through Playwright — navigate, click, type, extract, evaluate JS.
• Parallel isolation. Each agent gets its own BrowserContext; sub-agents can drive separate tabs concurrently without stepping on each other.
• Accessibility-tree observation. Token-efficient snapshots instead of screenshots — deterministic, cheap, no vision model needed.
• Full action replay. Every session is recorded with inputs, outputs, and DOM snapshots; replay with retry and output verification.

Ghost OS

• Screen awareness. Agents see what's on your display through a vision sidecar.
• MCP-native. ghost_parse_screen and ghost_annotate are exposed as tools the main agent can pick up.
• Pre-flight health checks. Sidecar auto-boots before the first call; status visible in MCP settings.

Channels

Connect your agent to your apps. Not through webhooks — as native integrations.

Channel	Setup	What works
WhatsApp	Scan a QR code	Messages, voice notes, attachments
Telegram	Paste a bot token	Messages, voice bubbles, interactive buttons
Slack	Socket Mode	Messages, files, native UI elements, threads
Discord	Paste a bot token	Messages, threads, buttons, attachments

Voice notes are transcribed automatically. Pair with the scheduler for cron-based delivery.

Features

Voice Input	Instant transcription, background grammar polish, concurrent recordings, per-recording cursor memory, dedicated transcriber model
Voice & Avatar	STT (cloud/local, 32 languages), TTS with voice cloning, 3D avatar with lip-sync
Images	Local or cloud generation, reference images, ComfyUI workflows as agent tools
Video	Images → MP4 with transitions and overlays
Design Workspace	Generate UI components with AI, live sandbox preview, edit-in-place via patches
Chromium Workspace	Agent-driven embedded browser (Playwright), parallel session isolation, accessibility-tree snapshots, full action replay
Ghost OS	Agents can see your screen via a vision sidecar
Folder Sync	Sync folders directly to agents; native FSEvents on macOS; replaces the old Knowledge Base
Deep Research	Multi-pass web search with cited writeups
Memory	Surfaces suggestions after conversations; you approve what sticks
Scheduler	Cron, interval, or one-time tasks — delivered to any channel or kept in chat
Skills	Reusable agent instructions. 37+ built-in, create your own from the UI
Plugins	Bundle skills and tools together. Install from GitHub or a URL
Workflows	Multi-agent delegation with parallel sub-agents and auto-delivered results
MCP	Connect external services as agent tools
Hooks	Run custom logic before or after any agent action
Workspace Styles	Classic Sidebar or Browser Tabs layout
Themes	8 color themes, light/dark, 50 wallpapers (20 live), rich text prompt editor

Providers

Use any combination, or go fully local with no API keys. Selene has its own architecture and it uses other clis like an api.

Provider	Models
Anthropic	Claude (Opus 4.7, Sonnet, Haiku) via API or Agent SDK
OpenAI	GPT-5 family + Codex
OpenRouter	Claude, Gemini, Grok, DeepSeek, and hundreds more
Ollama	Any local model; dynamic thinking detection
vLLM	Self-hosted inference
Kimi / Moonshot	K2.5, K2.6, K2.7 Code; OAuth device-flow sign-in
Minimax	Multiple variants
Antigravity	Free tier via Google OAuth

Download

macOS. Signed DMG, drag to Applications. Windows. Signed installer or portable build.

One download, no prerequisites. Selene bundles everything: Electron 41 (Chrome 146 / Node 24), local model support, browser engine, platform tools. The app is larger than usual because it ships what other tools make you install separately.

Grab the latest build on the Releases page.

For Developers

Setup

npm install
npm run electron:dev

Build

# Windows
npm run electron:dist:win:nosign

# macOS
npm run electron:dist:mac:nosign

Troubleshooting

• Native module errors: npm run electron:rebuild-native (rebuilds against the bundled Electron ABI)
• Embeddings mismatch: reindex from Settings
• MCP ENOENT: reinstall from latest DMG/installer

Thanks

Built on open-source. See THANKS.md.