开源AI工作流：Personal Reasoning Engine

macOS Prerequisites (Apple Silicon)

macOS Prerequisites (Intel + eGPU)

Note: Intel Macs run the Web GUI only (CLI requires Apple Silicon). PRE detects eGPU connections at runtime (30-second polling) — you can plug in an eGPU after PRE is already running and it will be detected automatically.

Windows Prerequisites

Linux Prerequisites

Optional: evolution-data-server (calendar/contacts/reminders via GNOME EDS), xdotool + scrot (desktop automation, X11), espeak-ng (TTS), xclip (clipboard).

What the installer does

All three installers follow the same pattern: check prerequisites, install missing dependencies, pull the Gemma 4 model from Ollama, auto-select quantization based on available memory, create ~/.pre/ data directories, size the context window, set up the Web GUI, and optionally configure auto-start. Everything stays local — no accounts, no cloud setup, no telemetry. The install is idempotent: running it again updates without breaking your existing data.

PRE detects your project, loads memories, and drops you into an interactive prompt:

╔══════════════════════════════════════════════════╗
║  Personal Reasoning Engine (PRE) 2.0            ║
║  Gemma 4 26B-A4B                                ║
╚══════════════════════════════════════════════════╝
  Server:  http://localhost:11434
  Web GUI: http://localhost:7749
  Project: my-project  /Users/you/my-project
  Channel: #general
  Memory:  3 entries loaded
  Type /help for commands

my-project #general>

---

Installation

macOS Install

git clone https://github.com/sunkencity999/pre.git
cd pre
./install.sh

The installer handles everything: system validation, Ollama, model pull, binary compilation, web GUI dependencies, terminal-notifier, ComfyUI (optional), data directories, RAM-based context window sizing, MCP auto-setup for Claude/Codex/Antigravity, model pre-warming, fast tier model creation (pre-gemma4-fast), inference backend selection (Ollama vs llama.cpp), and optional auto-start at login.

Inference backend: The installer auto-selects the best backend for your hardware. NVIDIA GPU users (via eGPU) get llama.cpp (Flash Attention + GBNF grammar constraints). Apple Silicon users get Ollama (simpler model management, similar performance). You can switch backends anytime from Settings.

./install.sh --yes   # Non-interactive — accepts all defaults

Or install manually:

```bash

Build PRE CLI + Telegram bot

make pre telegram

Windows Install

git clone https://github.com/sunkencity999/pre.git
cd pre

Easiest: Double-click install.cmd in the pre folder. This handles execution policy automatically.

From a terminal:

powershell -ExecutionPolicy Bypass -File install.ps1
powershell -ExecutionPolicy Bypass -File install.ps1 -Yes   # Non-interactive

The installer checks system requirements, installs Ollama and Node.js via winget, pulls the model, creates the fast tier model (pre-gemma4-fast), creates ~/.pre/ directories, auto-sizes the context window based on RAM, offers llama.cpp as the default backend for NVIDIA GPUs (with Ollama as fallback for embeddings), configures Ollama environment variables, and optionally enables auto-start at login.

Note: The Windows installer sets up the Web GUI only. The CLI engine (pre.m) is an Objective-C application that requires macOS. The Telegram bot is included in the Web GUI and works on all platforms.

Linux Install

git clone https://github.com/sunkencity999/pre.git
cd pre
./install-linux.sh

The installer detects your distro and package manager (apt/dnf/pacman), checks NVIDIA VRAM via nvidia-smi, installs Ollama, pulls the model with VRAM-aware quant selection, creates the fast tier model (pre-gemma4-fast), offers llama.cpp as the default backend for NVIDIA GPUs, installs Node.js dependencies, auto-sizes the context window, and optionally installs voice tools, GNOME PIM integration (evolution-data-server), desktop automation tools (xdotool/scrot), and systemd autostart.

Note: Linux runs the Web GUI only. The CLI engine requires macOS. Native app integration (calendar, contacts, reminders) requires GNOME with Evolution Data Server.

Quick Start

Optional: clickable cron notifications

brew install terminal-notifier

Enterprise Integrations

16 services in one interface — Jira, Confluence, SharePoint, Smartsheet, Slack, Linear, Zoom, Figma, Asana, Dynamics 365, Gmail, Google Drive, Google Docs, GitHub, Telegram, Brave Search, and Wolfram Alpha. Search Jira, cross-reference Confluence, pull a file from SharePoint, and post a summary to Slack — in one conversation.

Native app integrations (zero config) — Mail, Calendar, Contacts, Reminders, Notes, and Spotlight work immediately with whatever accounts you've configured. On macOS, uses Mail.app, Calendar.app, Contacts.app, Reminders.app, and Notes.app via AppleScript/EventKit. On Windows, uses Outlook COM for mail, calendar, contacts, and tasks, plus local markdown notes. Spotlight uses Windows Search on Windows. No API keys, no OAuth on either platform.

Head-to-Head Comparison