Omni-NLI exposes natural language inference capabilities through both a traditional REST API and the MCP interface used by AI agents. The REST API lets you send a JSON payload with a premise and hypothesis and receive a predicted label (entailment, contradiction, or neutral) along with a confidence score and model/backend information. The MCP interface provides a programmatic way for agents to request inferences, retrieve results, and integrate NLI checks into automated workflows or reasoning pipelines. The server is designed to be self-hostable, scalable, and configurable, with built-in caching to speed up repeated inferences and support for different backends such as Hugging Face models, Ollama, and OpenRouter. To start you’ll run the Omni-NLI server locally, after which you can query either the REST endpoints or the MCP endpoints depending on your integration needs.

Prerequisites:

• Python 3.10 or newer
• pip (comes with Python)
• Optional: access to HuggingFace models (internet access) or local model backends

Installation steps:

1. Create a virtual environment (recommended) python -m venv venv source venv/bin/activate # On Windows use: venv\Scripts\activate

2. Install the Omni-NLI package (with optional HuggingFace extras) pip install omni-nli[huggingface]

3. Run the Omni-NLI server

   This uses the module entry point configured by the package

   python -m omni_nli

4. (Optional) Run tests or verify installation

   Example: curl the REST API once the server is running at http://127.0.0.1:8000/api/v1/nli/evaluate

Notes:

• If you want a different backend or additional features, install with the appropriate extras or consult the documentation for configuration options.

Tips and common considerations:

• The MCP interface is intended for agent-based integrations; use the MCP endpoints to perform inferences as part of an automated reasoning workflow.
• The server provides caching to improve throughput for repeated inferences; configure cache size and expiry if needed via the documentation options.
• Backend models can range from public HuggingFace models to private/gated deployments; ensure proper authentication and access as required by your chosen backend.
• If you run into port conflicts or need to expose the MCP REST endpoints, check the configuration documentation for host/port settings.
• For production deployments, consider containerizing the service (Docker) and tuning resources based on your workload; refer to the official docs for deployment patterns.