SmartBrain SmartBrainA personal AI assistant that runs entirely on your computer. Your knowledge, your models, and your credentials never leave your hardware — encrypted under a passphrase only you hold.
brew install --cask securecloudgroup/tap/smartbrain
Detecting your system…
Runs entirely on your machine. No accounts, no telemetry — the only SmartBrain-operated service is an optional node for phone access that relays encrypted connections, serves the phone app, and keeps no records.
Your knowledge base, notes, and secrets live in a local database, encrypted at rest under your passphrase. Real documents are welcome — a several-hundred-page PDF is fine.
The assistant can read freely, but anything that changes data or reaches out waits for your approval — and every attempt is audited.
Group knowledge into vaults and share them — sealed for someone you trust, or published publicly so anyone can subscribe by URL and get signed updates.
Bring your own API keys, or run fully local models with Ollama or Apple MLX. A built-in gateway routes between them.
Every item here links to the page of the manual that covers it, so you can read the detail — including the limits — before you install anything.
Ask in plain language. The assistant can search your knowledge, read or summarize a whole document, save a note, add a task, look something up on the web, or draft an email. Answers stream as they are written, and you can stop, copy, regenerate or listen to one — and resend any message of yours with a tap on Retry.
Chat & toolsTap the mic or hold Space and speak; it stops when you pause, and the words appear as you say them. Nothing to set up — a built-in engine transcribes on your own machine, and audio never leaves it, even from your phone. Have replies read aloud, or switch on Conversation mode and just talk, with a wake word of your choosing.
VoiceDrop in PDFs, Word, PowerPoint, Excel, HTML, Markdown, CSV, JSON and plain text, paste a URL, or follow a website's RSS feed and let it fill a vault on its own. Search by keyword, by meaning, or both at once. Every result names its source and, where the file has them, the place in it — a deck cites its slide, a spreadsheet its sheet, a PDF its page — and opens the document at the passage that matched, highlighted.
KnowledgeTools are tiered. Reading runs on its own. Anything that writes or reaches out waits for you — approve it once, or tell it to stop asking about that one tool. Sending an email or deleting something asks every time, with a second confirmation, and can never be pre-approved. Every attempt goes into an audit log you can read.
Activity & approvalsGroup documents into a vault to scope a search — or to share it. Hand someone a sealed, encrypted file and the key separately, or publish an open one and let anyone subscribe by URL. Updates are signed with your key, and a subscriber's copy will not accept a change signed by a different one. You host the file; SmartBrain hosts nothing.
VaultsTasks with due dates, times and priorities, grouped into Today, This week and Later. The assistant can propose one; you approve it before it lands.
PlannerRun a prompt once, or hourly, daily or weekly — “every morning, summarize my open tasks.” Output lands in your chat and on the schedule's own page. Schedules run only while the app is unlocked, and anything risky inside a run parks for your approval rather than acting alone.
SchedulesOptional, and Gmail only. You connect it with your own Google OAuth client, and SmartBrain asks for two scopes: read and send. Nothing else — no archiving, no labels, no deleting. The assistant drafts; you approve the send.
EmailDuckDuckGo works with no key at all. You can switch to your own Brave or Tavily key, or point at a SearXNG instance you run. Searching is a reviewed action, so nothing goes out until you allow it.
Web searchYour name, its name, standing instructions and a list of remembered facts — encrypted, and added to every conversation. It is one readable list, so when you wonder why it keeps doing something, you can look, and delete the line.
MemoryRun models on your own machine with Ollama or Apple MLX, or bring your own keys for OpenAI, Anthropic or Google. A gateway on your machine routes each job — chat, reasoning, embeddings, summarizing — to whichever you picked for it.
Connect a modelAn estimate of what your cloud models have cost, per model, over a range you choose. Local models show as free. It is computed on your machine from your own token counts — no usage data leaves it.
Usage & costPair a phone once, then use SmartBrain on Wi-Fi or cellular. There is no port to forward and nothing listening for inbound connections. The link is direct where the network allows it; where it does not, a relay passes it along without being able to read it. Off until you pair a device.
Remote accessPoint Claude Desktop, Cursor or another MCP client at SmartBrain and let it search and read your knowledge base. Read-only, behind a token you mint yourself, and off until you do.
MCPExport everything as JSON, or take a backup you can restore later — your content stays encrypted under your own key inside it. Your Recovery Key and Emergency Kit are issued once at setup and are the only way back in if you forget your passphrase. There is no reset, because nobody else holds a copy.
Backup & recoveryIf you switch it on, SmartBrain scores its own recent work from telemetry that never leaves your machine. At most it learns one short preference at a time — it shows up in Memory marked “(learned)”, it is measured, and it is reverted on its own if it does not help. A second switch lets it steer how it handles a kind of request, and answers it steered say so. It does not change its own code.
Self-improvementThe app makes no calls of its own: no accounts, no analytics, no telemetry. Things go out when you ask for them — a cloud model you configured, a web search, Gmail, a vault you subscribed to — and the desktop app checks GitHub for its own updates. All of it is written down, along with the limits this design does not pretend to solve.
Privacy & securityPick your system. These install with no security warning — the app is delivered through package managers that vouch for the download, so you never hit an “unidentified developer” wall.
One command in the Terminal app. Installs to Applications, opens on first click — no security prompt, and on Apple Silicon nothing else to install.
No Homebrew yet? If brew says “command not found”, install it first from brew.sh (one command), then run this.
brew install --cask securecloudgroup/tap/smartbrain
In Terminal or PowerShell — no SmartScreen prompt, and nothing else to install.
No Scoop yet? If scoop isn’t recognised, install it first from scoop.sh (one command), then run this.
scoop bucket add securecloudgroup https://github.com/SecureCloudGroup/scoop-bucket; scoop install securecloudgroup/smartbrain
Download the install script, read it, run it — it verifies the release’s signature, installs per-user with no root, and adds SmartBrain to your app menu. Add --headless on a server for a systemd service. x86_64; arm and Alpine run the Docker stack instead.
curl -fsSLO https://raw.githubusercontent.com/SecureCloudGroup/SmartBrain_3000/main/installer/install-linux.sh && sh install-linux.sh
You run that command once. After the first launch SmartBrain looks for its own updates and
installs them when you say so — there is no package-manager command to run again. Prefer a
direct download? Grab the
latest release.
On Apple Silicon Macs, Windows, and x86_64 Linux it needs nothing else — it brings its own
runtime. An Intel Mac installs the same app with the same command and runs it in
Docker; containers are equally at home on
Linux (compose/docker-compose.release.yml in the repo). Docker Desktop, Colima,
OrbStack or Engine all work.
Full detail in Getting started.
From install to your first answer in a few minutes.
One command above. First launch downloads the app and opens it in your browser.
It encrypts everything on your machine. Save your Emergency Kit — it's the only way back in.
Add an API key or a local model (MLX or Ollama), then start building your private knowledge base.
Step by step, with screenshots, in First run and Your first 5 minutes.
The whole manual is public, and it is the same text the app shows on its own Help page.
Also worth reading: the changelog, the security policy, and the license.