neuraldefend-mcp

# neuraldefend-mcp `neuraldefend-mcp` 2.0.0 is a secure, local Model Context Protocol (MCP) server for NeuroVerify image and video authenticity analysis. It gives an MCP client two tools while keeping API credentials out of tool arguments, restricting which local media can be read, and returning a bounded integration contract instead of the Python SDK's raw response. Use the results as decision-support signals. A low-risk result is not proof that media is authentic, and a high-risk result should trigger human review rather than automatic rejection. ## How it works The MCP client starts `neuraldefend-mcp` as a child process and communicates with it over standard input and standard output: ```text MCP client └─ local stdio (MCP JSON-RPC) └─ neuraldefend-mcp ├─ validates and opens an allowlisted local file once └─ streams that open handle over TLS through the NeuroVerify Python SDK └─ NeuroVerify Detection API ``` There is no local HTTP listener. Standard output is reserved for the MCP protocol; safe startup errors are written to standard error. The server does not return media bytes to the MCP client and does not load an entire video into memory. It validates the opened file handle and uploads that same bounded handle, reducing path substitution and time-of-check/time-of-use risk. ## Requirements - Python 3.10 or newer. - A NeuralDefend API key authorized for the requested image or video analysis. - An MCP client that can launch a local stdio server. - Network access to the selected NeuroVerify API environment. - Read access to dedicated local directories containing only media authorized for upload. ## Installation Install the published 2.0.0 release: ```bash python -m pip install "neuraldefend-mcp==2.0.0" ``` The package installs the `neuraldefend-mcp` command. The equivalent module entry point is: ```bash python -m neuraldefend_mcp ``` Both entry points run the same stdio-only server. Starting either one direc…