llmgps is a self-hostable multi-LLM chat workspace.
- Modern Interface: Clean, responsive, and easy to use.
- Debate Mode: Advanced multi-model reasoning through structured disagreement.
- GPS Mode: Multi-model consensus and synthesis.
- Mainstream Presets: Out-of-the-box support for major AI providers.
- Ollama Integration: Full support for local model deployment.
- Encrypted Storage: History and keys saved in an encrypted SQLite database.
- Search & Proxy: Built-in web search support and proxy configuration.
Choose up to 5 models to tackle complex queries:
- Initial Phase: All models receive the prompt simultaneously.
- Consensus Check: A synthesizer evaluates if the models broadly agree.
- Debate Rounds: If there is friction, models review each other's points and refine their logic.
- Search Intervention: If disagreement persists, the system triggers web search to provide fresh grounding data.
- Synthesis: After a maximum of 2 rounds, the synthesizer generates a final, verified answer.
Choose up to 5 models. The system merges their independent responses into one cohesive final answer, highlighting key differences or specific insights from each model.
Pull and run:
docker pull ghcr.io/leodenglovescode/llmgps:main
docker run -d \
-p 3000:3000 \
-e LLMGPS_DATA_KEY=your-secret-key \
-v llmgps-data:/data \
--name llmgps \
ghcr.io/leodenglovescode/llmgps:mainAccess: http://localhost:3000
Set
LLMGPS_DATA_KEYto any strong secret string. The/datavolume persists your history and settings.
- Install:
npm install - Run:
npm run dev - Access:
http://localhost:3000
llmgps uses an encrypted SQLite database for all sensitive information.
- Encryption Key: Set via
LLMGPS_DATA_KEY. - Key Handling: In local dev, a
.llmgps-data.keyis auto-generated. In Docker, a key must be manually provided. - Persistence: Mount the
/datadirectory to keep your history and keys safe.
- Start Ollama.
- In llmgps Settings, enable Ollama and set your base URL (e.g.,
http://127.0.0.1:11434). - Select your local models directly in the GPS/Debate routing.
- Logic & Idea: @leodenglovescode
- Assistance: GitHub Copilot
Contributions are welcome! If you are familiar with the architecture, feel free to open a PR.