h2oGPT
h2oai/h2ogptReviewed by AI
Checked by RepoPicks Radar on
What is h2oGPT?
A chat program you run yourself: you upload documents (PDF, Word, Excel, images, audio, code, web pages) into a private database and ask questions or request summaries, with the answers coming from a model running on your own machine or from an online service you choose. It also does speech to text, reads answers aloud, generates images and can stand in for the OpenAI API. It runs on Linux, Windows and macOS, best through Docker, and wants an NVIDIA card for the heavier parts. The repository is archived, so nothing more will be fixed here.
h2oGPT is free and open source (Apache-2.0) and can replace ChatGPT, NotebookLM, and TypingMind.
How easy is it to install h2oGPT?
D · terminal, several tools or a specific graphics card · Windows
- Windows D
- macOS D
- Linux D
All three are a D. The README says "Docker is recommended for Linux, Windows, and MAC for full capabilities", which means Docker plus the NVIDIA runtime and a long build; the plain Linux script has the same capabilities but installs many Python pieces, and the README notes the "Windows and MAC scripts have less capabilities than using Docker". On top of that you download model files yourself and the releases contain no ready program.
Coming soon: install in 1 click with the RepoPicks app.
h2oGPT on GitHub
- Stars
- 11,924on Nov 10, 2026
- Downloads
- 595all releases
- Growth
- —no gain in the last reading
Stars over time
Facts
- Runs on
- Windows, macOS, Linux
- You need to install
- Python, Docker, NVIDIA graphics card
- To build from source
- Python, Docker, NVIDIA graphics card
- Disk space
- 950 MB–4.3 GB · estimate
- Accounts or keys
- OpenAI, Azure OpenAI, Anthropic, Google, MistralAI, Groq, Replicate or Together.ai, only if you choose an online model instead of a local one
- Cost
- Free · Apache 2.0, free to run. The company sells a separate enterprise product: the README's second heading is "Try Enterprise Version for Free" linking to Enterprise h2oGPTe. If you point it at a paid online model instead of a local one, you pay that service.
- License
- Apache-2.0 (Permissive)
- Can replace
- ChatGPT Plus, NotebookLM, TypingMind
- Last change on GitHub
- Oct 9, 2025
h2oGPT risk signals (7)
- Attentionstale releases
Last release on 2024-06-04 (886 days ago)
automatic check
- AttentionDependencies with known vulnerabilities
Pinned dependencies with known vulnerabilities (OSV): gradio: GHSA-26jh-r8g2-6fpr, GHSA-279j-x4gx-hfrh, GHSA-39mp-8hj3-5c49; torch: GHSA-3749-ghw9-m3mg, GHSA-53q9-r3pm-6pq6, GHSA-63cw-57p8-fm3p; torch: GHSA-3749-ghw9-m3mg, GHSA-53q9-r3pm-6pq6, GHSA-63cw-57p8-fm3p
automatic check
- Attentionsome packages come from the company bucket
requirements.txt installs two packages straight from the company's storage instead of the normal Python index: "InstructorEmbedding @ https://h2o-release.s3.amazonaws.com/h2ogpt/InstructorEmbedding-1.0.1-py3-none-any.whl" and "sentence_transformers_old @ https://h2o-release.s3.amazonaws.com/h2ogpt/sentence_transformers_old-2.2.2-py3-none-any.whl". Whatever those addresses serve at install time runs on your machine, and the install breaks if they are taken down.
AI review
- Attentionweb interface published without a password by default
The compose files publish the interface ("ports: - '7860:7860'") and the Dockerfile exposes 7860, 8888, 5000, 5002 and 5004, while the login feature is optional ("Authentication in the UI by user/password via Native or Google OAuth") and no password is set in these files. On a machine others can reach, your documents and the chat would be open to them.
AI review
- InfoRepository archived
The repository is archived (read-only, no longer maintained)
automatic check
- Infooptional extras pull gpl like packages
The install lists "reqs_optional/requirements_optional_langchain.gpllike.txt" among the requirement files pulled by setup.py. The project's own code is Apache 2.0, but those optional packages carry stricter licences, which matters if you plan to build something commercial on top.
AI review
- Infoold pins that will not move
The requirements freeze 2024-era versions ("gradio==4.44.0", "torch==2.2.1", "spacy==3.7.5", "huggingface_hub==0.25.2") and the repository is read-only, so the versions flagged with known security problems cannot be raised here; anyone needing a fix has to change them in their own copy.
AI review
Evidence is shown in its original language.
What Radar could not check about h2oGPT
- Whether the install still completes today, since the pinned 2024 versions have to agree with current Python, CUDA and Hugging Face packages.
- Whether the Windows and macOS scripts still work, and which features are missing there compared with Docker.
- Real disk space and video memory once you download models, embeddings and the image or voice extras.
- How good the answers over your own documents are, and how much the paid online models would cost.
- Whether the demo and enterprise addresses in the README are still up.