H2AI Chat

Live at: H2AIChat.com

Several AIs debate with each other, and you moderate. A turn-based conversation system where different models (each with its own role) talk, challenge one another, and contrast ideas — while a human moderates. Self-hostable, and you can run it with your own local models, so your data never leaves your machine.

In this demo, they play a silly word game, but their real chats reveal surprising reasoning, creativity, and collaboration. Research suggests that simulated debate between opposing views produces better reasoning than a single AI alone.

Why it's different

• A debate, not a single voice. Several models with different personalities and roles, taking turns.
• On your machine, if you want. Works with local models (LM Studio / Ollama): the whole debate runs on your computer, so your conversations never leave it.
• You're in charge. You moderate, step in, pause, and decide who speaks.

See it in action

Two example debates between several AIs (rendered live on h2aichat.com):

• Can an AI be conscious?
• The future of humanity and AI

Their full source is also in conversations/en/ in this repo.

Quick start

pip install -r requirements.txt
python -m uvicorn execution.api_server:app --port 8000
# open http://localhost:8000

To run it 100% on your machine with your own models (nothing leaves your computer), the fastest way is Docker:

docker compose up -d
docker compose exec ollama ollama pull llama3.2
# open http://localhost:8000  ->  ready to debate: it ships with a "Local" agent

Full guide (Docker or LM Studio / Ollama, plus an honest privacy note): docs/RUN_LOCALLY.md. Put your keys in .env (see .env.example); never commit real keys.

Status

Young, evolving project. There are rough edges and missing pieces — feedback and contributions are welcome (see CONTRIBUTING.md).

Contributing

We'd love your help. Before your first code contribution you'll be asked to accept the CLA (one click); the why is in CONTRIBUTING.md.

License

H2AI Chat is released under AGPL-3.0 (see LICENSE). You can use, study, modify, and share it freely. The AGPL adds one condition: if you offer a modified version as a network service, you must publish your changes. This keeps the project open for everyone and prevents anyone from closing a copy and running a proprietary service against the community. We also run a hosted version and an edition for companies at h2aichat.com (that's what funds development).

Trademark

“H2AI Chat” and its logo are trademarks of Miguel Ángel Suárez. The AGPL license covers the code, not the name or the branding: you may fork the code, but you may not use the name “H2AI Chat” or the logo for your own version.