MyGPT is a privacy-first, open-source Retrieval-Augmented Generation (RAG) platform with real-time confidence metrics and in-document source highlighting. Ask questions across your research papers, clinical protocols, and document collections—with verifiable citations and zero third-party API costs, running entirely on your personal device or institutional infrastructure.
A modular RAG stack designed for enterprise, research, and personal use, prioritizing security and factual accuracy.
Run MyGPT on your personal device or private infrastructure so your library and conversations stay under your control.
Retrieved context, source highlighting, and confidence metrics help you inspect where an answer comes from and decide when to verify it.
MyGPT calculates real-time metrics for every query: Question Relevance Score (QRS), Answer Relevance Score (ARS), and Hallucination Index (HI).
Use Ollama to run supported open models locally. Choose the model that fits your hardware, quality needs, and privacy requirements.
Cross-platform Electron desktop preview for macOS, Windows, and Linux. Standalone window management and status indicators while macOS notarization is in progress.
Connect your web application or desktop preview to remote GPU workstations, institutional HPC servers, or Azure cloud deployments. Offload inference workloads while keeping personal laptops lightweight.
A modular workspace engineered for researchers, clinical teams, and data analysts.
Available as both a responsive browser application and a native Electron desktop preview (macOS, Windows, Linux). Simultaneously inspect questions with real-time confidence metrics, navigate library documents, and view cited in-document PDF highlights.
Seamlessly expose private document context to Claude Desktop, IDEs, and other MCP-compliant autonomous agent pipelines.
Directly connect your personal or shared Zotero libraries to instantly synchronize collections of papers without manual file copying.
Customize token lengths, overlap margins, Chroma DB vector metrics, and BM25 hybrid indexing tailored to your specific literature domain.
Switch between Llama 3.1, Gemma 2, DeepSeek-R1, Mistral, or edge models with customizable temperature and top-p sampling on the fly.
See exactly how the local pipeline process resolves queries step-by-step from raw question to certified answer.
MyGPT
Ask a question about a paper or a topic from your publication library. We will try to answer it using the GPT models.
© 2026
Set up MyGPT on your system using our prebuilt Docker images and script controls.
The official and recommended way to install and run MyGPT on your personal computer or workstation. Runs securely with Docker Desktop and Ollama, and opens directly in your browser at http://localhost:3000 without desktop signing restrictions.
Check the Desktop App (Preview) tab. Note that macOS developer signing is in progress, so the web application below is the recommended choice.
Requires at least 8 CPUs, 8 GB RAM (16 GB recommended for faster responses), and 10 GB disk space. Ensure Homebrew, Git, Docker Desktop, and Ollama are installed.
ollama pull llama3
ollama pull nomic-embed-textgit clone https://github.com/stjude/MyGPT.git
cd MyGPTcp .env_backend.example .env_backend
cp .env_frontend.example .env_frontendThese files configure your local container parameters securely without committing secrets.
cd installation/macOS
bash run_docker.shAlternatively, launch directly with Docker Compose from the root folder: docker compose up -d. Once started, open http://localhost:3000 in your browser.
See the Desktop App (Preview) tab for Windows setup and portable executables.
Ensure Docker Desktop (with WSL 2 backend) and Git are installed on your Windows PC.
ollama pull llama3
ollama pull nomic-embed-textgit clone https://github.com/stjude/MyGPT.git
cd MyGPT
copy .env_backend.example .env_backend
copy .env_frontend.example .env_frontendcd installation\windows
run_docker.batAccess the UI at http://localhost:3000.
See the Desktop App (Preview) tab for Linux AppImage and Debian packages.
Requires at least 8 CPUs, 8 GB RAM (16 GB recommended), and 10 GB disk space. Compatible with both CPU and NVIDIA GPU acceleration.
sudo systemctl start ollama
ollama pull llama3
ollama pull nomic-embed-textgit clone https://github.com/stjude/MyGPT.git
cd MyGPT
cp .env_backend.example .env_backend
cp .env_frontend.example .env_frontendcd installation/linux
sudo sh run_docker.shFor GPU systems with CUDA: docker compose -f docker-compose-gpu.yml up -d
Pre-packaged desktop builds for macOS, Windows, and Linux. Standalone desktop apps do not require Node.js, npm, or cloning the Git repository. For local use, simply have Docker Desktop and Ollama running in the background; or connect seamlessly to a remote GPU server or cloud deployment without running local Docker containers.
Run everything locally on your machine for guaranteed data privacy.
ollama pull llama3.2
ollama pull nomic-embed-texthttp://localhost:11434.Connect your desktop app to a shared GPU workstation, institutional HPC cluster, or Azure deployment.
https://mygpt.institution.edu:8000) and remote Ollama URL.If you are modifying the desktop app or building custom signed installers for your institution:
cd electron
npm install
npm run dev # Launch desktop app with hot reloading
npm run dist # Build installers for current operating systemSee the complete build and configuration documentation in electron/README.md.
Production server, multi-VM GPU clusters, and Azure cloud deployment guides are actively maintained in the official GitHub repository.
Direct links to specialized deployment guides on GitHub:
MyGPT was developed at St. Jude Children's Research Hospital to democratize privacy-first, verifiable literature exploration.
In text, cite this work as Patel et al. (2026) or (Patel et al., 2026).
@article{patel2026mygpt,
author = {Patel, Jaimin and Downing, Jude and Ferguson, Hugh and You, Thika and Malinverni, Duccio and Mathew D. A., Steve and Chen, Ines and Sluter, Madison and Moorefield, Beth and Parej, Katalin and Ragavan, Mukundan and Morris, Cindy and Keerthi, Dinesh and Becerra Armada, Desiree and Meszaros, Balint and Trivedi, Vikas and Alam, Shahinur and Woodard, Anthony and Alford, Dan and Pathak, Sagar and Li, Cai and Umeton, Renato and Rodriguez-Galindo, Carlos and Lam, Catherine G. and Gottschalk, Stephen and Kalodimos, Charalampos G. and Babu, M. Madan},
title = {Democratizing reliable knowledge-seeking with MyGPT: A Privacy-First, Open-Source Retrieval-Augmented Generation Platform},
journal = {Nature Communications},
year = {2026},
doi = {10.1038/s41467-026-78509-6},
note = {Early Access},
archive = {10.5281/zenodo.22165249}
}Comprehensive evaluation datasets, test queries, and scoring pipelines archived in the official MyGPT-evaluations repository.
Domain-specific question answering benchmark evaluated against gold-standard biomedical literature with curated reference PDFs.
View BioASQ on GitHubQuestion answering on PubMed abstracts assessing entity re-ranking and factual precision against verified ground-truth citations.
View PubMedQA on GitHubOpen-domain QA benchmark evaluating retrieval-augmented generation on full-text arXiv scientific research articles.
View Open-rag-bench on GitHubMultilingual evaluation over national and regional public health policy PDFs (French, Spanish, and English) testing retrieval without external leakage.
View Health Policies on GitHubBiomedical literature mining benchmark evaluating kinase assay information extraction from PubMed-linked papers with synonym resolution.
View Kinase Literature on GitHubThreshold calculation dataset establishing optimal Question and Answer Relevance cutoff boundaries across 5 local embedding models.
View Cutoff Calibration on GitHub