UFriend

English | 繁體中文

A Windows AI workspace designed primarily for ASUS UGen300 local computing.

UFriend brings email, recordings, image text, and everyday writing into one local workspace. Its primary hardware target is ASUS UGen300: the goal is to turn a dedicated USB AI accelerator into an everyday desktop assistant. Review the original material alongside each result, edit it, and decide how to use it. The interface supports English and Traditional Chinese, with light, dark, and system themes.

Version: 0.1.42 · Windows x64 · Electron 44

Releases · Documentation · User manual · Report an issue

Designed for ASUS UGen300

UGen300 is the main hardware platform UFriend is designed around. The intended experience pairs the accelerator with compatible models and a local service, so users can work with AI through familiar desktop workflows.

Design focus	Intended benefit
Local data processing	Process supported text, image, and speech tasks on the connected accelerator, keeping model inputs on the user's computer and device.
Low-power AI acceleration	Use a dedicated AI processor for supported inference workloads, with hardware power consumption suited to everyday desktop use.
Reduced host workload	Move supported model inference to the accelerator and its dedicated memory, while the PC handles the interface, files, and task coordination.
One practical workspace	Bring local AI into email organization, OCR, transcription, translation, and writing without requiring users to manage a separate interface for each workflow.

ASUS lists the UGen300 USB model with a Hailo-10H processor, 8 GB LPDDR4, up to 40 TOPS (INT4), and 2.5 W typical power consumption. These are the manufacturer's accelerator specifications; UFriend's model compatibility, latency, throughput, and total-system power still require hardware measurement. See the ASUS product page.

Current implementation: The desktop workflows are available through separately configured PC model services or cloud APIs. The UGen300 device backend and Hailo SDK integration are still pending; selecting the reserved UGen300 option does not yet run inference on the device. Device-pool timing simulations do not perform model inference. See the integration contract and next steps.

Local processing is already the default for text, OCR, and speech, and failed tasks do not automatically switch to the cloud. Gmail, external email images, and explicitly selected cloud AI still use network connections. See model sources and data flows.

Features

Workflow	Available functionality
Email	Import EML files or connect Gmail; read HTML mail, manage attachments and drafts, send mail, summarize and classify messages, and schedule daily organization.
Recordings	Import audio or video, choose a transcription language, queue or schedule jobs, jump to timestamps, and edit or export TXT and SRT transcripts.
Captioned video	Burn timed subtitles into a new MP4 with progress and cancellation, keeping the source video. The Windows release includes FFmpeg.
Image text	Import an image or read a clipboard image, run OCR, compare the source with editable text, and copy the result. Automatic clipboard capture is opt-in.
Writing	Stream chat responses, keep separate conversations, translate text, and polish writing by purpose and tone. Optional profile context and task prompts are configurable.
Workspace	Local persistence and recovery backups, data cleanup, diagnostics, and system tray operation.

AI results need review. Mail summaries currently use Traditional Chinese, independently of the interface language. Meeting tools provide transcription and subtitles; speaker diarization and automatic meeting summaries are not implemented.

Run the Windows release

1. Open GitHub Releases and download the UFriend Windows EXE attachment for Windows x64. Source archives and the separate FFmpeg developer package are not needed to run UFriend.
2. Save the EXE in a folder you want to keep. If another version is running, quit it from the system tray menu first.
3. Double-click the downloaded EXE. No installation wizard, Node.js, or manual extraction is needed.
4. For English, open 設定 → 一般 → 軟體介面語言, choose English, and save.

The launcher extracts the application into a separate temporary directory for each run. Workspace data and encrypted credentials remain in your Windows user data directory; they are not stored inside the downloaded EXE or its temporary extraction folder. Restarting from the system tray launches the original downloaded EXE. Keep that file in place until you fully quit the app.

The Windows release includes FFmpeg for captioned-video export and audio preparation for cloud speech transcription. No separate FFmpeg installation or 7-Zip is needed. Settings → General → Media runtime shows that the runtime is included. The release also provides its complete corresponding source, dependency sources, and build scripts as a separate download.

Older releases and builds made without FFmpeg provide these optional installation controls instead:

• Download and install media runtime downloads the pinned build from Gyan and verifies its hashes. An existing 7-Zip installation is required.
• Select extracted package opens a folder picker for an already extracted official package. It does not download anything or require 7-Zip.
• Media runtime instructions opens the runtime guide.

A runtime installed through the app is saved in your user data directory and remains available after restarting or replacing the EXE. Local Whisper uses its own independently installed runtime. AI model services and Gmail still require the separate configuration described below.

If you use a folder-based UFriend ZIP package instead, extract the entire ZIP, keep its DLLs, resources, and license files together, and run its UFriend.exe. UFriend folder packages built without FFmpeg also include Install media runtime.cmd; the runtime guide covers that alternative. The separate FFmpeg developer ZIP is not a UFriend app package.

Use the data-location control in Settings → General to find the workspace. Fully quit UFriend before backing it up or replacing application files.

Configure AI and Gmail

Local AI

The desktop repository contains HTTP adapters, not model weights or model-service source code. The following services are independent projects and are currently available by contacting the maintainer. Cloning this repository or installing a desktop release does not install them.

Service	Default endpoint	Current implementation
PC LLM API	http://127.0.0.1:8765	Qwen2.5-1.5B-Instruct Q4_K_M with llama.cpp CUDA
PC OCR API	http://127.0.0.1:8767	PP-OCRv5 mobile detection and recognition on CPU
PC Whisper API	http://127.0.0.1:8768	Multilingual Whisper Small with PyTorch CUDA; Base is optional

Install and start each required service using its own README. In Settings → Models and compute → Local models, select PC local services under Local backend, save the endpoints, and test the connections. Only loopback endpoints are supported for these services. UFriend does not install, start, or stop them. See the PC model guide for requirements.

Cloud AI

Configure your own compatible API endpoint, key, and text, vision, or speech model in the cloud settings. Text and OCR use the OpenAI-compatible Chat Completions format; speech requires Audio Transcriptions with segment timestamps. Compatibility varies by provider and model.

Text, OCR, and speech have separate defaults, and you can select a source for each task. A failed task does not automatically switch to another source. See model sources and data flows.

Gmail

Bring your own Google Desktop OAuth client configuration. Enable Gmail API in your Google Cloud project, configure its consent screen and permitted users, then import the downloaded Desktop client JSON in Settings → Email accounts → Advanced settings (developers). Follow the Gmail setup guide, then authorize your account in the system browser.

The repository and public release contain an empty OAuth template; no shared client or Gmail credentials are supplied. EML import and local drafts can be used without a Gmail connection. Connected mail currently supports Gmail only.

Build from source

Use Windows x64 with Node.js 24.x, npm, Windows PowerShell 5.1, and the .NET Framework 4.x C# compiler. Electron is installed through the lockfile. The desktop itself does not require Python or CUDA; separate model services have their own requirements.

git clone https://github.com/breeze0305/UFriend.git
cd UFriend
npm ci
npm run build:native

Prepare the pinned FFmpeg runtime and the UFRIEND_TEST_FFMPEG test path as described in the installation guide and FFmpeg runtime guide. Then run:

npm run check
npm test
npm start

After quitting the app, build the single-file Windows EXE:

npm run package:portable

This uses the pinned electron-builder 26.17.0 NSIS portable target. The first build needs network access to download its build tools. It creates and validates a folder package without FFmpeg or embedded OAuth settings, then writes UFriend-<version>.exe and its build receipt under local/releases/UFriend-<version>-portable-<timestamp>-<id>/. The verified folder package remains under dist/. New builds use this filename without packaging or runtime suffixes; already published assets retain their original names.

For the release format with FFmpeg already inside the single EXE, first prepare the pinned UFriend-built runtime, then run:

npm run package:portable -- --with-ffmpeg

This also creates UFriend-<version>.exe, in its own build directory. The build receipt records whether FFmpeg is included. A bundled runtime takes precedence over an older optional installation in user data. The executable includes the GPL license, third-party notices, and source information. Publish the matching complete source archive alongside the EXE; a successful binary build alone does not complete this step. See the runtime guide.

To create only the folder package for ZIP distribution, use npm run package -- --without-ffmpeg. For a local folder package using the prepared FFmpeg runtime, use npm run package. Packaging does not run the full test suite or publish a release. See the installation guide for isolated development data, UI smoke tests, prerequisites, and packaging details.

Architecture and data

flowchart LR
    UI[Electron renderer] --> Main[Main process and task control]
    Main --> Data[Local workspace and encrypted credentials]
    Main --> Gmail[Gmail API]
    Main --> Local[Independent local LLM / OCR / speech services]
    Main -. planned .-> UGen[UGen300 / Hailo local service]
    Main --> Cloud[Configured cloud API]
    Main --> FFmpeg[Local FFmpeg process]

Loading

• Mail, files, settings, and completed work are stored locally, normally under %APPDATA%\UFriend\workspace. Ordinary workspace files are not encrypted by UFriend. Credentials use Electron safeStorage, backed by the Windows user account. An app-installed media runtime is stored separately under %APPDATA%\UFriend\media-runtime\ffmpeg.
• Gmail operations connect to Google. External email images connect to their source websites. Local inference does not make every feature offline.
• Cloud tasks send the necessary input to the configured endpoint. Chat includes the current conversation history; enabled profile context is included only in the applicable text tasks. Provider retention and charges depend on the provider.
• Model output does not independently send mail, delete messages, create calendar events, or execute programs.

The dashed UGen300 path shows the planned hardware backend described in Designed for ASUS UGen300. Outlook, Yahoo, and calendar integrations are also outside the current supported scope.

Documentation and project files

Project documentation is available as separate English (.md) and Traditional Chinese (.zh-TW.md) files. Each documentation index lists both versions.

• User and developer documentation
• Model adapter contracts
• Windows clipboard component
• Contributing and security reporting

app/ contains the desktop application, scripts/ the build and verification tools, and tests/ the automated tests. Dependencies, release output, credentials, personal data, and model runtimes are excluded from source control.

License and third-party components

UFriend is licensed under the MIT License.

Third-party components retain their respective licenses. These include Electron, Lucide, DOMPurify, MailParser, Nodemailer, and OpenCC. FFmpeg is run as a separate process and has its own GPL license; UFriend's MIT license does not replace it. See third-party notices, the FFmpeg runtime and source information and Google asset attribution.