Gliner2

MCP server that exposes GLiNER2 entity extraction, classification, and JSON extraction as tools.

Quickstart

1. Install dependencies:

uv sync

2. Run the MCP server:

uv run gliner2-mcp

Environment configuration

This server is configured via environment variables (optionally loaded from .env).

If GLINER2_MCP_ENV_FILE is not set, the server will load .env from the working directory when it exists.

MCP tools

• extract_entities(labels, text=None, filename=None) returns a mapping of label to entity spans.
• classify_text(schema, text=None, filename=None) returns a mapping of field name to chosen label or list of labels.
• extract_json(schema, text=None, filename=None) returns a JSON object that matches the provided schema.

For all tools, provide exactly one of text or filename:

• text: inline input text.
• filename: UTF-8 text file path relative to the current working directory. Absolute paths and paths resolving outside the current working directory are rejected.

Legacy camelCase aliases are still available for compatibility: extractEntities, classifyText, extractJson.

Available Models

GLiNER2 family models on Hugging Face

Development

uv sync --dev
uv run ruff check .

Testing

Run the default test suite (unit + contract tests):

uv sync --extra dev
uv run --extra dev pytest -q

Run real-model integration tests (downloads/loads actual GLiNER2 weights):

GLINER2_MCP_RUN_INTEGRATION=1 uv run --extra dev pytest -q -m integration

Recommended pre-release check:

uv run --extra dev pytest -q
GLINER2_MCP_RUN_INTEGRATION=1 uv run --extra dev pytest -q -m integration

License

MIT. See LICENSE.