The desktop app is being prepared for release. Everything below describes how it will work; for now, use Neutropic in your browser — sign up and start, nothing to install.
What it adds
The desktop app is the same workspace as the web, plus a local helper that runs the file and code tools on your computer:
- Local folders — allow specific folders; the agent reads datasets, recordings and documents from them without uploading.
- Local code execution — data inspection, analysis and figure code runs on your machine; only observations and artifact metadata go to the server.
- Media pack — an optional download that enables audio/video analysis (transcription, facial action units, pose, rPPG) locally.
- Sign-in through your browser — the app opens the sign-in page in your default browser and returns via a
neutropic://link. - Local AI — run whole turns on a model installed on this computer (a bundled llama.cpp runtime, or your own Ollama / LM Studio server). Local turns never use managed tokens.
Everything else — cloud models, search, memory, billing — works exactly as on the web.
Install
Get it from the download page when it opens (macOS and Windows). On macOS, open the .dmg and drag Neutropic to Applications; on Windows, run the installer. The helper is bundled; the media pack is an optional download from Customize → Storage.
Set up local access
- Open Customize → Storage.
- Under Local folders, press Add folder and pick the folders the agent may read.
- If you plan to analyze audio or video, press Install media pack.
In a chat, refer to a local file by name — "analyze study3.csv in my Data folder" — or attach it with + as usual; attached files are read locally when the app is running.
Local AI
The desktop app can run a whole turn on a model that lives on your computer. This is desktop-only: in the browser and on mobile, turns run on cloud models.
- Open Customize → Local AI.
- Press Install runtime — llama.cpp is downloaded once. If you already run Ollama or LM Studio (or any OpenAI-compatible server), the panel detects it and you can use that instead, with nothing to download.
- Install a model from the recommended list. It is sized to your RAM: 8–16 GB suits a 4B-class model, 24–32 GB an 8B–20B-class, and 64 GB or more a 30B-class model.
- In a chat, open the Model menu and pick the model under On this device.
What to expect:
- Local turns never use managed tokens and work on every plan, including Free.
- Lite and Basic effort are available; Deep is not available on local models yet — a Deep turn runs as Basic with a notice.
- Local models are slower than the cloud. The first step of a turn can take a while while the prompt is processed; later steps are quicker. If the model fails (out of memory, app closed mid-turn), the chat offers Retry with a cloud model.
- In a sensitive project, everything — including routing and summaries — stays on the local model. In other projects, the Use cloud for routing & summaries (uses managed tokens) toggle in the panel can send those small helper calls to a fast cloud model; it is off by default.
Where your data goes. The model runs on your computer. Prompts and tool observations pass through Neutropic servers in memory to reach it, but no third-party AI provider receives your data. Local turns never use managed tokens. Your original files are still read only on this device, as with every desktop turn.
More on effort levels and the model menu: Effort levels & models.
Storage on this computer
The same page shows artifacts made on this computer and code-execution working copies, each with a Clear button. Clearing working copies never touches your original files.
Updates
The app checks for updates on launch and installs them with your approval. Versions are listed in the changelog under Desktop app. The desktop app is in its final release preparation — the changelog says when it is generally available.
Trouble? See Desktop app issues.