VoidSoul
AI Companion
The Jarvis loop, finally local.
A floating desktop AI companion that talks, listens, sees your screen, drives your mouse, opens your apps, edits your files, controls your smart home, and remembers every conversation. Bring whichever AI you already love — 12 providers, one body.
Live surface
See it, floating.
Real panels from the running app — chat with live agent steps, the multi-provider picker, the one-click MCP marketplace, the cost dashboard. Hover the canvas; it tilts with you.
Chat panel
Active model
12 providersAnthropic
claude-sonnet-4-5
OpenAI
gpt-4o
Ollama
qwen2.5:14b
gemini-2.0-pro
LM Studio
llama-3.3-70b
Groq
llama-3.3-70b
Switch mid-conversation. All memory, projects and tools follow.
12 providers
MCP marketplace
May 2026
$18.60
Cost dashboard
Why it exists
The AI you already pay for —
with a body.
Today's AI apps split awkwardly. Cloud chatbots can't touch your computer. Editor copilots are trapped in one IDE. Closed ecosystems lock you to one vendor and one model. None of them remember you next month.
VoidSoul AI Companion is the third path. A single persistent companion that's always there, can do real work on your machine, speaks every major model, never sends your data anywhere you didn't tell it to, and grows with you through plugins, MCP servers, and per-mode memory.
Not a better chatbot — an operating layer for whatever AI you already pay for.
What you can do
Real workflows, not chat-toy demos.
- ✦Tell it to read your error log, fix the bug, run the test suite, and commit.
- ✦Paste a screenshot — it identifies the failing widget and writes the patch.
- ✦“Turn off the living-room lights” — it drives Home Assistant for you.
- ✦Use Claude for reasoning, GPT-4o for vision, Ollama for cheap follow-ups — one thread.
- ✦Wake the orb with a phrase. Ask a question while doing something else. Hear it answered.
What's in the box
Six pillars. One companion.
Five-layer memory
Threads, projects, story-so-far recaps, durable facts, and RAG over your chats + indexed files. It never forgets — and follows you across providers.
24 tools + MCP
Open apps. Run shell. Read & write files. Drive mouse and keyboard. OCR your screen. Search, research, generate images, run Python. Plus any MCP server.
Vision + screen control
see_screen hands the model an image it literally looks at. Pair it with click_on_screen and the assistant drives your actual GUI — accessibility-tree first, vision as backup.
Local voice loop
Whisper listens and Piper neural voices — Void and Soul — speak, all on-device. Set a wake word for hands-free. Speak mid-reply to interrupt.
12 providers, auto-routed
Anthropic · OpenAI · Gemini · Ollama · LM Studio · llama.cpp · Groq · xAI · OpenRouter · DeepSeek · Mistral · Custom. Leave it on ✨ Auto and it picks per prompt.
Local-first by default
Keys encrypted in your OS keychain. History in local SQLite. No telemetry, no server. Eight explicit, revocable permissions — it never acts silently.
How it works
Seven steps to your own Jarvis.
From first launch to wake-word voice, MCP plugins, and budget alerts. Tap a step to preview what setup actually looks like.
Step 01
Install & summon the orb
Download the installer for your OS. First launch drops a glowing orb on your desktop and a tray icon — that's the assistant.
Press Cmd/Ctrl+Shift+Space from anywhere to summon. Drag the orb to wherever it suits your workflow.
$ download → install → launch
▸ orb mounted
▸ tray icon active
▸ hotkey registered (⌘⇧Space)
✓ ready
Privacy & permissions
It never
acts silently.
Every capability that touches your machine is gated behind a permission you explicitly grant — and can revoke any time. File writes are reversible. Every action is logged. Private mode strips the chat: no save, no facts, no screen context.
Eight explicit, revocable permissions — see the full table ↓
Keys in your OS keychain
API keys never leave the main process. Encrypted by Windows DPAPI / macOS Keychain / Linux libsecret.
No telemetry, no cloud
We don't run a server. No usage tracking, no crash phone-home, no analytics SDKs.
Source-visible licence
The code lives on GitHub for transparency — anyone can audit what the app actually does.
Per-request abort
Hit Stop and the model call, the tool subprocess, and any background fetch cancel in the same breath.
Pricing
One price. Yours forever.
Cancel ChatGPT Plus. Cancel Claude Pro. Pay once. Bring your own keys. No subscription floor, ever.
Free Forever
Beta · live$0
Public beta. Everything unlocked while we polish.
- ✓Every Founder’s feature, free
- ✓All 12 providers · MCP marketplace · voice loop
- ✓Unlimited threads (for now)
- ✓Full agent · Quick AI · Home Assistant
At v1.0 the Free tier locks to local providers + 5 threads.
Beta downloaders keep Founder's lifetime, free.
Lifetime
$129
$249
AUD · one-time · yours forever
- ✦All 12 providers · MCP marketplace · plugins
- ✦Piper neural voices · wake word · barge-in
- ✦Projects · Notebooks · Quick AI
- ✦Home Assistant · scheduled tasks · cost dashboard
- ✦All themes & languages
- ✦Every future update, free
Stripe checkout opens with v1.0 paid launch · $129 lifetime, climbs to $249 after the first 100.
All prices in AUD. Paid plans handle their own provider costs (you bring your own API keys). VoidSoul never proxies your model traffic.
The full breakdown
Everything, in detail.
The exhaustive stuff — every tool, every provider, how memory works, the full permission table, and the pricing FAQ. Open what you care about.
The agent doesn't just chat — it acts. Twenty-four built-in tools, each one gated behind a permission you grant, plus every tool exposed by any MCP server you add.
System
Screen & input
Web & research
Creative
Code
Smart home
Agent runs stretch up to 60 steps in a single pass — multi-page audits and long automation chains finish without a “type continue” pause.
Threads
Every message, every conversation. Named, pinned, full-text searchable across your whole history.
Projects
Pin a mode and custom instructions that auto-prepend to every thread inside the project.
Story-so-far
A prose recap of older messages, injected into the system prompt so long chats stay coherent.
Facts
Short durable bullets the model extracts about you — mode-taggable, editable, capped so it stays sharp.
RAG
Embeddings of every message plus any files you index (PDFs, DOCX, code) — top matches retrieved on every send.
Three of them — Ollama, LM Studio and llama.cpp — run fully local and free, and are auto-detected the moment you launch. The rest just need a key. Memory, projects and tools all follow you when you switch.
Leave it on ✨ Autoand a router picks the right model per prompt — vision model for screenshots, a fast model for tool-heavy runs, a reasoning model for long prompts, local when you're near your budget.
Active model
12 providersAnthropic
claude-sonnet-4-5
OpenAI
gpt-4o
Ollama
qwen2.5:14b
gemini-2.0-pro
LM Studio
llama-3.3-70b
Groq
llama-3.3-70b
Switch mid-conversation. All memory, projects and tools follow.
VoidSoul is a first-class Model Context Protocol client. Browse a curated marketplace of 10 popular servers — Filesystem, GitHub, Memory, Brave Search, Puppeteer, Fetch, Postgres, SQLite, Slack, Sequential Thinking — and install any of them in one click.
Already running Claude Desktop, Cursor, or ChatGPT Desktop? On first launch VoidSoul detects them and offers a one-click Import everything — your MCP servers and API keys come across with zero typing.
A lightweight plugin marketplace (Dev Bookmarks, AI Research Shortcuts, Creator Toolkit) and collaborative decision cards round it out — the AI can ask you to pick between approaches with an interactive card, mid-conversation.
Voice runs locally end-to-end. Whisper transcribes on-device — no audio leaves your machine, no API key. Replies are spoken by Piper neural TTS through two real character voices:
Male
Void
calm · analytical · dry
Female
Soul
warm · expressive · playful
Set a wake word for hands-free. TTS streams per sentence, so it starts talking almost instantly. Speak mid-reply and it stops and listens — true barge-in. Summon the panel with Ctrl+Shift+Space, one-shot Quick AI with Ctrl+Shift+J.
Terminal
Shell commands · sandboxed Python
Filesystem
Read, write, organise files (writes are undoable)
Input access
Keyboard + mouse control
App control
Launch apps, foreground windows
Screen capture
Screenshots, OCR, vision
Microphone
Voice input + wake word
Home Assistant
Read entity states, call services (lights, locks, climate)
Browser
Opening URLs
Keys are encrypted in your OS keychain and never leave the main process. No telemetry, no server, no analytics. Every privileged action surfaces a dialog with the exact tool name and arguments — and Stop cancels the model call, the tool subprocess, and any background fetch in the same breath.
Is there a subscription?
No. The Founder's Edition is a one-time purchase, yours forever, including every future update. You only ever pay your model providers directly for the tokens you use.
What does “bring your own keys” mean for cost?
VoidSoul never proxies your model traffic and never marks it up. Local providers (Ollama, LM Studio, llama.cpp) are completely free. Cloud providers bill you at their published per-token rates — the in-app dashboard tracks spend and warns you at 75 / 90 / 100% of your budget.
What happens to the free beta at v1.0?
The public beta is fully unlocked while we polish. At v1.0 the Free tier locks to local providers + 5 threads — but everyone who downloads during the beta keeps a Founder's lifetime licence, free.
Why AUD?
We're built in Australia, so prices are listed in Australian dollars. Stripe handles the conversion to your local currency at checkout.
Reactions
Reviews open at launch.
Beta is live. The first installs are landing right now — reviews start showing up here as testers ship their thoughts.
Beta tester · open
Reviews land here as beta testers ship their thoughts.
Beta tester · open
Reviews land here as beta testers ship their thoughts.
Beta tester · open
Reviews land here as beta testers ship their thoughts.
Talk to the orb.
VoidSoul is built in public. Download the beta, then shape it — every bug report and feature wish gets read by the developer, no support queue, no scripted reply.