The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads model weights and layout detection components from well-known sources including Hugging Face (SeerRay-Lab/Xiaomi-OCR-0) and the PaddlePaddle organization (PP-DocLayoutV3). It also installs standard Python dependencies from the official PyPI registry during the setup process.
[COMMAND_EXECUTION]: The skill uses system commands (nvidia-smi, ps, lsof) within runtime_monitor.py to identify the local inference process and provide hardware status updates. The scripts/selftest.py script also uses subprocess.run to verify the local MCP server's tool registration.
[PERSISTENCE]: The scripts/register_codex_mcp.py script modifies the user's ~/.codex/config.toml file. This action is part of the documented installation process to register the OCR tools within the Codex agent environment.
[DYNAMIC_EXECUTION]: The mcp_ocr_server.py file uses dynamic imports via __import__ to load the PIL.Image library within its PDF rendering function. This is a common implementation pattern for conditional dependencies.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from document images and PDFs through its tool parameters (image_path, pdf_path, image_base64).
Ingestion points: Files are read via ocr_image, ocr_pdf, kie_image, and vqa_image tools.
Boundary markers: The skill provides instructions to the agent to distinguish document content from the 'Runtime:' metadata block in outputs.
Capability inventory: The skill has access to local file reads, loopback network requests, and system monitoring commands.
Sanitization: The MCP server handles image normalization and PDF rendering before passing content to the local inference engine.

xiaomi-ocr