GLiNER Extraction MCP

Extract anything. Redact everything. A hosted MCP server for zero-shot entity extraction and PII redaction, powered by GLiNER — no training, no label taxonomy, no per-field setup.

Point any MCP-capable agent (Claude, ChatGPT, Cursor, custom agents) at this server and it gets structured extraction and compliance-grade redaction as native tool calls.

Tools

extract_entities

Pass text plus any label set you can phrase in words:

{
  "text": "Per our call, ACME will deliver 40 units by March 3rd. Contact Sarah Lin (s.lin@acme.io).",
  "labels": ["organization", "quantity", "deadline", "person", "email"]
}

Returns each match with exact character spans, label, and confidence. Invoice fields, contract clauses, medication dosages, job titles, product SKUs — if you can name it, it extracts it. That's the zero-shot difference: classic NER models know ~18 fixed types; GLiNER matches your labels at inference time.

redact_pii

Two detection layers, one merged result:

• Model layer — contextual PII a regex can't see: names, addresses, organizations, dates of birth
• Pattern layer — structured PII a model can miss: emails, phone numbers, SSNs, credit cards (Luhn-validated), IP addresses

Three redaction styles: label → [PERSON], mask → █████, remove → deleted. Pass custom labels to redact domain-specific fields (e.g. ["patient name", "diagnosis"]) instead of the default PII set.

Privacy: text is processed in memory only — never stored, never logged.

Usage (MCP)

{
  "mcpServers": {
    "gliner-extraction": {
      "url": "https://<actor-standby-url>/mcp"
    }
  }
}

Pricing

Pay per event, metered by document size (1 unit = 10,000 characters, minimum 1):

A 3-page contract ≈ 1 unit. Cheap enough to run on every document in a pipeline — that's the point.

Batch mode

Run the Actor directly with { "mode": "redact", "texts": [...], "labels": [...] } to process a corpus into the dataset.

Model

Default: urchade/gliner_small-v2.1 (Apache 2.0), baked into the image — no cold-start download. The build arg GLINER_MODEL swaps in any GLiNER checkpoint (e.g. urchade/gliner_multi-v2.1 for multilingual).

Limits

• Max 200,000 characters per call, 25 labels per request.
• Detection is probabilistic: for regulated use, treat redaction output as a strong first pass with human review, not a guarantee.