Gemina FileTag

What Gemina FileTag MCP server does

Gemina FileTag MCP server exposes three related capabilities through one hosted MCP endpoint: FileTag processing, structured extraction, and document intelligence. It accepts PDFs and common image formats, including PNG, JPEG, GIF, WebP, HEIC/HEIF, and AVIF, up to 50 MB.

FileTag identifies document metadata, produces six filename patterns, and returns a short-lived URL for a copy with metadata embedded in the file. The free tier includes 1,500 FileTag tags per month. Extraction is a separate, paid capability for invoice fields, invoice line items, OCR, custom templates, and supported legacy Hebrew extraction types.

Indexed documents can be searched with structured filters, semantic similarity, or hybrid search. Aggregations provide sums, averages, and counts grouped by fields such as vendor, currency, document type, or month.

How it works

The service is hosted at https://api.gemina.co/api/v1/mcp/; there is no local daemon to install. The transport is Streamable HTTP. Clients can authenticate through an OAuth 2.1 browser sign-in, or use an API key for headless environments such as scripts and servers.

For a file upload, call files_create_upload to reserve a pre-signed PUT slot. The response contains a file ID, upload URL, required headers, and a suggested next call. Upload the bytes to that URL, then call tag_file for FileTag processing or extract_document when the upload was created for extraction. A public HTTPS resource can instead be passed to tag_url, keeping its bytes outside the model context.

Extraction may complete asynchronously. Use get_extraction_result with the returned correlation ID to poll it, or use the listing and retrieval tools to inspect stored extraction records. Documents already held by Gemina can receive additional extraction types without another upload.

Tools and capabilities

• files_create_upload: Reserve an upload slot and receive PUT instructions.
• tag_file: Return metadata, six filename patterns, and an enriched-file URL.
• tag_url: Fetch and tag a public HTTPS URL server-side.
• extract_document: Run invoice, OCR, custom-template, or supported legacy extraction types.
• get_extraction_result, list_extractions, get_extraction, and get_document: Poll or retrieve extraction data.
• add_document_extractions: Run more extraction types on a stored document, either waiting for values or filing the work for later.
• submit_extraction_feedback: Send corrected field values back to Gemina.
• query_documents: Search indexed documents with structured, semantic, or hybrid modes.
• aggregate_documents: Calculate grouped sums, averages, minimums, maximums, and counts.
• index_document: Reindex a document after corrections or for backfilling.

Setup and configuration

Point an MCP-compatible client at https://api.gemina.co/api/v1/mcp/. With OAuth enabled, the client discovers Gemina's authorization service and opens a browser sign-in; no local process or copied key is required. Create a free Gemina account if needed.

For headless use, configure the client to send the Gemina API key as an X-API-Key header. The repository provides client-specific installation examples, while the hosted endpoint handles the server runtime. Clients that need to download enriched files may also need network access to the URL returned by tag_file.

Limitations and notes

FileTag supports a free monthly allowance, while structured extraction and document intelligence are paid capabilities. Extraction and indexing behavior depends on the selected extraction type and account credits. Plain OCR is not automatically submitted for document indexing, and a document may be skipped when it has no extractable fields or indexing credits are unavailable.

Enriched-file URLs are temporary, so download the result while the URL is valid. The API-key method is intended for headless operation; interactive clients can generally use the OAuth sign-in flow instead. The MCP server itself is operated as a hosted service, and this repository contains its installation material and examples rather than a local server implementation.