The off-ramp from the AI silos.

Personal AI on hardware you own. Private by default, free forever.

Get it on GitHub Install with setup
Local-first No telemetry AGPL licensed Active development

Personal AI is being decided right now.

Their cloud, their interests, your data. The off-ramp is local, on hardware you own, with memory that doesn't leak. Somebody had to build it. This is it.

Move personal AI from cloud silos to your machine A locked vendor cloud on the left connects by an off-ramp route to a local Augmentum machine on the right. vendor silo account, policy, meter off-ramp bring memory home your machine local memory, local models

One private layer, many shapes.

Pick the mode that matches what you're doing, or let Augmentum route the request.

Everyday conversation

Fast chat with inline helpers and memory that's actually yours.

Structured analysis

Break complex requests into search, fetch, verify, and response steps with tools in the loop.

Long-form storytelling

Three-layer memory keeps characters and plot consistent across thousands of turns.

Real deliverables

Tell it the goal — get back a .docx, .pptx, .xlsx, or EPUB.

Code with you, not at you

Plan / act / test loop in a sandboxed workspace. Every change reviewable.

Across every surface you use.

One conversation, four screens, no copy-paste.

  • Voice on the phone kicks a flow; the brief lands on the laptop.
  • Opt-in game and cast surfaces can turn a phone into a controller and a TV into the shared screen.
  • With Fabric enabled, a tablet can borrow the tower's GPU for the hard prompts on your configured network.
  • Claude Desktop, Cursor, Cline reach into the same memory via MCP.

What's bundled.

The core stack runs locally. GPU, Fabric, and extra media services are opt-in overlays when your hardware supports them.

Inference engine

Bundled llama.cpp. Auto-discovers your GGUF models, connects to Ollama/LM Studio, or downloads from the catalog.

Sandboxed Python

Containerized interpreter for tool execution. Code can't break out.

Image generation

Optional NVIDIA GPU overlay for SD15 / SDXL / FLUX locally; cloud image providers can be configured explicitly.

Voice

Moonshine STT plus local TTS options such as Kokoro and Pocket; optional overlays add more engines.

Web search

Bundled SearXNG queries upstream search engines for you without requiring a third-party API key.

Knowledge packs

Import Wikipedia, MDWiki, Stack Exchange, DevDocs-style packs for offline hybrid retrieval.

MCP server

Claude Desktop, Cursor, Cline — anything MCP-speaking — use your substrate.

Fabric peering

Opt-in peering across boxes you own. Seven capability kinds can route peer-to-peer, with no Augmentum-hosted middleman.

Multi-user auth

Argon2id, opaque sessions, fail-closed middleware. Every memory scoped to its owner.

Clone it. Run setup. Start local.

The setup wizard writes .env and .augmentum.conf, detects CPU vs NVIDIA GPU paths, and asks whether to start Augmentum when it finishes. It binds to 127.0.0.1 until you say otherwise.

macOS / Linux

git clone https://github.com/AugmentumHQ/Augmentum augmentum
cd augmentum
./setup.sh

Windows PowerShell

git clone https://github.com/AugmentumHQ/Augmentum augmentum
cd augmentum
.\setup.bat

Choose CPU or NVIDIA GPU in setup. If you accept the start prompt, open https://localhost:6443 and accept the local certificate warning once; plain HTTP is also available at http://localhost:6100/ui.

Optional: put it on the TV.

experimental

Cast any surface — chat, avatar, audio, video, comics, or a game — to a TV, with a phone as the controller. The receiver runs in any browser: open https://<server-ip>:6443/ui/cast-receiver/ on the TV's browser, then scan the QR it shows with a phone that's logged into Augmentum. No install needed.

Want a dedicated, always-on TV client instead? Build the native Android TV receiver and sideload it with adb — the most experimental piece of the stack.

Sideload the Android TV APK with adb

1 · Build the APK needs Java 17 + Android SDK; no prebuilt APK yet, first build takes 3–5 min

cd augmentum/cast/android-tv-receiver
./gradlew assembleDebug

2 · Enable developer mode on the TV

Settings → System → About → click "Android TV OS build" seven times, then enable Network debugging under Developer Options. Note the TV's LAN IP from Settings → Network.

3 · Connect and install accept the debugging prompt on the TV once

adb connect 192.168.1.42:5555
adb -s 192.168.1.42:5555 install -r app/build/outputs/apk/debug/app-debug.apk

4 · First launch

Open the Augmentum Receiver app, enter your server's URL (https://<server-ip>:6443), then pair by scanning the on-screen QR with a phone that's logged into Augmentum. Pair codes expire after ~2 minutes — a fresh QR appears if one lapses. The app auto-starts on TV reboot after that.

Caveats: some TVs reject the self-signed certificate — accept it in the TV's browser first. Updates are a manual rebuild + reinstall. Full walkthrough: Android TV Receiver guide.

Three things I won't do.

If any of these change, stop supporting the project.

Won't sell you out

No ads, no telemetry without consent, no selling user data.

Won't take features away

If it works today, it works tomorrow on the same terms.

Won't sell out to investors

No funding with strings attached. Grows through community or stays small.

Help keep it free.

Use it daily. Share it. Contribute code, docs, packs, characters, translations.

Support options One-time tip

Same product whether you give nothing, five dollars, or five hundred.