space ocr

A field extraction API with built-in verification. Send a document photo with a list of field names — one POST returns typed JSON. No training, no processor setup, no SDK. Every value carries the box / quad it was read from (coordinates come from the OCR pass, not the language model), and values that fail the cross-check come back on an explicit review list — a work list, not a bare confidence score. Measured side by side against Mistral Document AI on the same photos, same schema, same grading script: field accuracy 91.1% vs 70.0%, and 16 of 463 values moved across three runs vs 100 (8 cases over 7 documents · 463 fields · 3 runs each · August 2026). Three endpoints, one verification contract: POST /ocr/fields for named fields, /ocr/markdown for layout-preserving Markdown, /ocr/text for reading-order text. Reads Japanese, Korean, Chinese and English. Optional document store: folders and sheets with server-side queries (GET /view) and an MCP server for AI agents. $0.05 per scan, failed scans never charged, 100 pages free every month, no card required. Rate limit 60 requests/min per key; observed latency p50 7.2 s per image (not an SLA).

Use the remote MCP URL on this page in your MCP client (Claude, Cursor, VS Code, and more). Copy the endpoint if your client needs a manual entry, then complete provider sign-in when your client connects — tools appear after authentication.

This endpoint requires provider sign-in, so MCPBundles could not list tools anonymously during the last probe. After you connect and authenticate in your MCP client, your session should receive the live tool list from the provider.

space ocr may require signing in to the provider before tools can run. Connect through MCPBundles or your MCP client and complete any provider login when prompted.