Private Context AI

A private desktop chat app for local AI models.
Your prompts go to Ollama on this computer and nowhere else.

Features

• Local chat with any model you have pulled into Ollama, streamed as it is written.
• Encrypted history. Saved chats are kept in an encrypted database on your disk.
• Temporary chats. Their text stays in the app's memory and is never written to disk. Their request metadata still shows in Activity.
• Projects with shared instructions and an automatic memory of the project's other chats, which you can view, edit and turn off.
• Compare models. Ask several models the same question side by side, then continue with one.
• Charts and thinking, both off by default: let a model draw a bar, line or pie chart, or show its reasoning.
• Markdown replies with code highlighting, copy buttons and Copy as Markdown.
• Activity lists every request the app made. Models shows each model's speed.
• Search across your chats, plus System, Light and Dark themes.

How it works

flowchart LR
    subgraph app["Private Context AI"]
        UI["React UI<br/>no network access"]
        Core["Rust core"]
        DB[("Encrypted database<br/>chats · audit log · settings")]
    end
    Keychain["OS keychain"]
    Ollama["Ollama<br/>127.0.0.1:11434"]
    Internet(("Internet"))

    UI <-->|"Tauri IPC"| Core
    Core <--> DB
    Keychain -.->|"database key"| Core
    Core <-->|"HTTP, loopback only"| Ollama
    Ollama -.->|"model downloads<br/>Ollama's own traffic"| Internet

Loading

The window has no network access of its own. Only the Rust core talks to Ollama, and only on this computer. Ollama itself is a separate program: it downloads models and may check for updates, but that traffic is Ollama's, not this app's.

Security at a glance

• No network but Ollama on loopback. A strict CSP blocks the window from the network. Rust refuses non-loopback hosts, proxies and redirects, and it is built without TLS, so it cannot speak HTTPS at all.
• Encrypted at rest. Chats live in a SQLCipher database. Its key is kept in the OS keychain and never reaches the window.
• Every request is audited. A row is written to the audit log before each request goes out; if that fails, the request is refused.
• Logs hold no content. Only metadata such as request kinds, counts and timings.
• The model cannot act. It gets at most one display-only chart tool. Replies render without raw HTML, and links and images are inert.

What it does not protect against: Ollama itself, malware running as you, the clipboard and the screen. Read docs/security.md for the full model and threat notes.

Quick start

You need Node.js 26 (exact version in .nvmrc), pnpm 12 (exact version in packageManager), Rust (rustup) and Ollama. Node.js 26 no longer bundles Corepack, so install it once with npm install -g corepack && corepack enable. Start Ollama first: open the Ollama app, or run ollama serve in its own terminal.

pnpm install
ollama pull llama3.2             # any model works
pnpm --filter desktop dev        # open the app
pnpm --filter desktop build:app  # build installers

Platform setup (WebView2, webkit2gtk, SQLCipher build tools, WSL) and every repo command are in docs/development.md.

Keyboard

Key	In the chat
Enter	Send the message
Shift+Enter	New line
Esc	Stop the reply being generated

Esc works wherever focus is in the chat, except in the sidebar's rename field and in a message being edited, where it cancels the rename or the edit. In the message editor, Enter saves and resends.

Your data

Saved chats, the audit log and settings live in one encrypted database in the app data folder, with its key in the OS keychain. Log files hold metadata only. docs/data.md lists the folders and files, and explains how to wipe everything.

Docs

• Security model and threat notes
• Where your data lives
• Development, building and commands