impossibleG/impossible-ocr

A self-hosted OCR server that turns PNG and JPEG images into structured text using local PP-OCRv5 models through HTTP, gRPC and MCP.

About impossible-ocr

Impossible OCR is a local-first OCR server. Version 0.1 is deliberately scoped to synchronous OCR of static PNG or JPEG raster images with the curated PP-OCRv5 mobile detector and English mobile recognizer bundle. HTTP, unary gRPC, and MCP all use the same bounded pipeline.

The server starts on loopback by default. Liveness is available immediately, while readiness stays false and OCR requests return a stable unavailable error unless the optional ONNX backend is enabled, its exact local bundle is installed, its integrity-bound native runtime is admitted, and every session passes warm-up. This fail-closed behavior is intentional.

- impossible-ocr-domain: validated raster input, options, errors, geometry, and canonical results. - impossible-ocr-protocol: transport-neutral HTTP, gRPC, and MCP boundary contracts. - impossible-ocr-pipeline: backend lifecycle and synchronous pipeline traits. - impossible-ocr-onnx: explicit model lifecycle, bounded graph admission, and optional CPU ONNX Runtime adapter. - impossible-ocr-server: CLI plus bounded HTTP control plane and transport extension modules. - vendor/impossible-server: provisional reviewed source snapshot of the neutral server foundation.

- Product contract - Model admission boundary - Native ONNX Runtime packaging - Deferred work and cross-project lessons

The backlog records possible follow-up work only. It does not block the current v0.1 scope and does not commit the project to dates or features.

From the project README.

Maintaining this server?

Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:

[![On MCP Radar](https://mcp.liqiwa.com/badge.svg)](https://mcp.liqiwa.com/s/impossibleG--impossible-ocr.html)

Related MCP servers