Your Own AI

Private AI on your machine. No one in control but you.

Your Own AI is a free desktop app for macOS, Windows, and Linux that runs AI models privately on your device. Your AIs, their personalities, their memories of you, and every conversation live on your machine - with an offline model, nothing you type ever leaves it. And because the app is open source, that claim is verifiable, not promised.

Download · Models · Website · Documentation

• Runs locally, on the machine you already have. Download open models (GGUF) from a curated catalog with a fit check for your hardware, and chat with no account or internet for local use. Models larger than your graphics card still run: a mixture-of-experts model keeps its rarely used parts in main memory and the rest on the card, and the catalog carries one for every size of machine, from 12 GB laptops to big-memory workstations. GPU acceleration out of the box (Vulkan on Linux and Windows, Metal on macOS), an optional one-click CUDA engine for NVIDIA cards, and models can live on any drive. Every model, offline and online, has its own page at yourownai.net/models.
• Memory that knows you as a whole. Each AI has one memory - what you have given it and what it has learned with you - and everything your AIs know about you lives in one place. On top of it, your AIs keep a short summary of who you are, written on your own device by a local model and rewritten as things change. Every fact is yours to read, edit, or forget, traceable to the conversation it came from; if the summary is wrong, fix the fact and it rewrites itself. None of it leaves your machine.
• Routing that measures instead of assumes. Auto modes pick the model for each question by how models actually run on your computer - speed measured from real use, room for the turn, what the question needs - not a spec sheet. Online models follow the preferences you set. Every reply says which model answered and why, health questions stay on your device by policy, and the mode name is the consent: "Auto - Offline Only" never touches the internet.
• Characters that are yours. Author AIs from 18 personalities with their own portraits, memories, and knowledge - for personal or work use - then share them as single-file packs, cryptographically signed so anyone importing one can verify who made it. Eight free characters are on the characters page.
• Projects: agentic coding in chat. Open a folder and your AI reads files, proposes edits as diffs, and runs commands - with a permission level you choose, and every step recorded. Powered by Your Own AI Build, a free add-on downloaded in one click.
• Documents and images, with proof. Attach files (scanned paper included, via on-device OCR) and images (read by a vision model on your device). A long document gets the room it needs on your machine when your machine can give it, and an explicit offer when only an online model can. Check any answer claim-by-claim against the exact wording in the source with Verify sources.
• Records you can prove. Conversations are written into tamper-evident, cryptographically signed records (built on Holochain), private and encrypted on your device, from the moment you send - a stopped reply is kept and marked. Export any conversation, and optionally sign it with your Flowsta identity so anyone can verify it at flowsta.com.
• Backed up, recoverable. With a Flowsta Vault connected, everything backs up automatically - each conversation as its own compressed object, so no session is too large to protect - and restores to a new device.
• Online when you want it. Frontier models and web search are an optional paid service, signed in with your Flowsta identity, through a relay that strips your identity and never stores or logs your messages. Never required, never default - and everything local stays free if you cancel.
• An engine for other apps. A local OpenAI-compatible endpoint serves your AIs - personality, memory, and records included - to editors, agent frameworks, and scripts. Or connect your own OpenAI-compatible server and its models join the picker.

• macOS 10.15 or later (Apple Silicon or Intel), Windows 10/11 (64-bit), or a 64-bit Linux desktop (.deb or .rpm).
• 8 GB of memory runs the small models; 16 GB and a graphics card with 6 GB or more opens up most of the catalog; 32 GB carries the large mixture-of-experts models with the graphics card doing part of the work. The models pages check any model against your machine before you download it, and the app grades every model the same way once installed.

• Tauri v2 (Rust backend)
• Qwik (TypeScript frontend)
• Tailwind CSS
• llama.cpp (local inference)
• Holochain (signed, encrypted conversation records)
• Your Own AI Build (the agent behind Projects)

npm install
./download-llama-binaries.sh # fetches the inference engine sidecars
npm run tauri dev

Rust tests run with cargo test --lib in src-tauri/. Building the full app also bundles Holochain sidecar binaries and pdfium; see .github/workflows/build-release.yml for the complete recipe. Releases are tagged vX.Y.Z (betas vX.Y.Z-beta.N) and their notes come from CHANGELOG.md.

Found a problem? Open an issue with the diagnostics file from Settings - it carries the app log and your hardware summary, and nothing personal. Betas go to a small group of testers first; every stable release is what came out of their reports.

AGPL-3.0-or-later - see LICENSE and THIRD_PARTY_NOTICES.