Characters, roleplay, branching stories, and a private assistant — powered by AI models that run on your own GPU, not the cloud. One download sizes itself to your hardware, then works fully offline: no setup, no account, nothing phoning home.
A ~35-minute deep dive through the whole app — first-run setup, building characters, the emotion-reactive live portraits, the routed expert assistant, and the privacy model that keeps it all on your machine.
Not a chat box bolted onto a server. Aphelion is one app where characters, stories, dialogue trees, and an expert assistant live side by side — all powered by models running on your own machine. Whether you're drafting a novel without a chaperone, querying a folder of your own research, or just want a capable AI that answers to no one — it's all one window.
Every prompt, model, and conversation runs and lives on your PC. Your data never leaves it — online or off.
Install one program and start talking — no command line, no separate tools to download and wire together.
Aphelion reads your GPU and VRAM and loads the best model that runs fast — no config files, no guesswork.
Characters and group chats, long-form stories, branching dialogue trees, and expert assistants — all in one place.
No account. No telemetry. No phone-home. The lock icon means exactly what it says.
Free and MIT-licensed. Read the source, change it, build on it. Nothing hidden.
On first run, Aphelion checks your graphics card and downloads a model sized to fit — so it runs fast instead of crawling. Bigger card, bigger and smarter model. Switch any time, or bring your own.
A compact model like Gemma 3 4B. Runs on modest GPUs — even CPU in a pinch.
A strong all-rounder like Gemma 3 12B for mid-range cards.
Flagship models — Gemma 3 27B or SuperGemma4 26B — for high-VRAM GPUs.
Not sure how much VRAM you have? You don't need to know. VRAM is just your graphics card's memory (more of it = a bigger, smarter model) — and Aphelion checks yours on first run and picks one that fits. The numbers above are only if you're curious.
Uncensored models are available behind an opt-in warning, and you can point Aphelion at your own GGUF file. Whatever you pick, it runs entirely on your machine.
Some "AI" tools are autonomous agents — deliberately given permission to run commands, browse the web, and change files on your computer (think AutoGPT, OpenClaw, and computer-use bots). That power is also the risk. Aphelion is the opposite: the model only ever speaks text — it has no autonomy and no reach into your system.
The model itself never touches your disk — when you grant a folder, the app reads the files and feeds the model plain text. It can write a villain's monologue; it can't go rummaging through your hard drive on its own.
Yes — completely. MIT open source, no account, no trial, no upsell.
No, never. The AI runs on your own computer. Disconnect from the internet and it keeps working.
There is no telemetry by default — nothing is collected, nothing is sent. If you want to help squash bugs, Settings → Crash reports is an opt-in switch (off until you flip it): when the app hits an unexpected error, an anonymous report goes to sentry.io containing the error, stack trace, app version, and OS — never your chats, characters, prompts, or files. The reporting code isn't even loaded until you opt in, and you can turn it off any time. All told, Aphelion touches the internet only when you tell it to: downloading a model, checking for updates, or crash reports if you opted in.
No. The model only generates text — it has no autonomy. The app reads files only from folders you explicitly point it at, and saves documents you ask it to create; it never runs commands, browses the web, or goes through your files on its own. Unlike autonomous agents, it has no free rein over your system. See "It's a model, not an agent" above.
The AI's "brain" — a file the app downloads once. Setup picks a good one for your hardware; you can change it in Settings.
No — it just means the app isn't code-signed yet, and that certificate is a recurring monthly cost, so while Aphelion is free I haven't bought one. Click More info → Run anyway (the source is all public). If enough folks chip in, I can sign it and the warning goes away.
Windows 10/11, macOS (Apple Silicon), or Linux (x64). A graphics card (GPU) is recommended — more VRAM means a bigger, smarter model — but setup fits a model to whatever you have. macOS and Linux builds are brand-new (beta) and less battle-tested than Windows — please report anything odd.
Yes — behind an opt-in warning. Whatever the AI writes is generated by the model you choose, not by Aphelion. Use at your own discretion.
Settings → Apps → Aphelion → Uninstall. It even asks whether to also delete the downloaded model and your chats, so you can reclaim the disk space.
Full history on the GitHub releases page.
Aphelion is free and open source. It's also unsigned — the only reason Windows shows that "unknown publisher" warning — because a code-signing certificate is a recurring monthly cost, and I haven't paid for one. If donations start coming in, that's the first thing they'll go toward. They also help keep Aphelion free: the more they offset what it costs to build and ship, the less chance I'd ever have to charge for it. No pressure — if it's useful to you, anything helps.
No pressure, ever — the app stays free either way. Donations just decide how fast the rough edges (like that signing warning) get smoothed out.