OpenAssist is a local-first AI assistant for real-time work. It listens to system audio, reads screen context when needed, and routes questions through fast AI providers or local models.

The goal is simple: capture context from the environment you are already working in, understand the actual question, and return a useful answer quickly.

• AI coding assistant for live debugging, architecture questions, and code review.
• Meeting copilot for system-audio transcription and follow-up answers.
• Interview assistant for technical and behavioral question practice.
• Screen-aware assistant for OCR, docs, terminals, and browser workflows.
• Local knowledge assistant with RAG over project files and documents.

• Turns spoken audio into actionable questions with Auto Mode.
• Captures screen/OCR context for code, docs, slides, terminals, and forms.
• Routes answers through Groq, Gemini, Cerebras, Together, Ollama, and other providers.
• Uses local RAG and semantic cache for project or knowledge-base context.
• Keeps session context so follow-up questions make sense.
• Supports local Whisper fallback when cloud transcription is unavailable.
• Runs as a desktop app with hotkeys, settings, and an optional overlay UI.

• Python 3.11+
• Windows 10/11 recommended
• API key for at least one cloud provider, or Ollama for local inference

Fastest path, one command:

bash setup.sh

That script creates the virtual environment, installs dependencies, prepares local folders, and sets up Ollama if it is missing.

Manual install if you want to do it step by step:

git clone <repo-url>
cd openassist

python -m venv venv
.\venv\Scripts\activate
pip install -r requirements.txt

Copy .env.example to .env or add keys from the in-app setup wizard.

Common providers:

• Groq: fast text generation and Whisper transcription
• Gemini: text and vision fallback
• Cerebras / Together: low-latency cloud alternatives
• Ollama: local models, no API key

You can also edit config.yaml directly during development. Local config and runtime data are ignored by git.

.\run.bat

or:

python main.py

On Linux/macOS, use:

bash run.sh

For system audio on Windows, OpenAssist prefers:

1. WASAPI loopback
2. Virtual audio cable
3. Stereo Mix, if explicitly enabled

Microphone fallback is opt-in because it is not a reliable substitute for meeting/system audio capture.

Drop files into knowledge/documents/ and the app can index them for retrieval. Supported content includes text, markdown, code, Q&A-style files, and PDFs where extraction is available.

You can also add documents with:

python main.py --add-docs path\to\docs

Run the app:

.\run.bat

Run tests:

python -m pytest

Run Auto Mode benchmark:

python benchmarks/auto_mode_benchmark.py --dir tests/fixtures/auto_ground_truth --out benchmarks/auto_mode_full.json

Run a focused test file:

python -m pytest tests/test_text_utils.py -q

ai/          Provider routing, prompts, cache, memory, RAG, intent logic
capture/     Audio capture, speech transcription, screen/OCR capture
core/        App lifecycle, config, hotkeys, session state
knowledge/   Local documents and ingestion helpers
modes/       Mode profiles and behavior tuning
ui/          Desktop UI, settings, overlay, standby screen
utils/       Text cleanup, crypto, platform helpers, telemetry
tests/       Regression and behavior tests
benchmarks/  Fixture-driven latency and quality benchmarks

• Auto Mode is the main real-time voice path.
• Generated benchmark JSON, logs, local cache, and learned runtime data are ignored.
• The app warms critical models before enabling Start Session.
• Cloud services can still vary in latency, so local fallback paths stay available.

MIT