CRC-LNM Multimodal Research Assistant MCP

This MCP server provides a six-tool, research-assistance workflow for allowlisted, deidentified CRC-LNM cases. It accepts only precomputed 1409-dimensional CT features, 768-dimensional pathology features, and four clinical values. It does not accept raw imaging files, file paths, or external feature vectors.

ModelScope STDIO Deployment (Quick Start)

Step 1: Select Service Type

Select "STDIO" (NOT "Streamable HTTP")

Step 2: Fill These Fields Separately

Step 3: Set Environment Variable

No environment variable or local path is required. The explicit CPU PyTorch index keeps the package request portable across hosted Linux workers.

Enter these as separate arguments. Do not paste a combined shell command into the command field.

Step 4: Deploy

Click deploy and wait for list_tools to complete with 6 tools.

Verification Order

1. Build and inspect the wheel, then run the console entry point from an unrelated working directory.
2. Publish the verified wheel to PyPI and start it with the exact uvx command above.
3. Let ModelScope complete list_tools, then manually test each required tools.
4. Obtain the ModelScope URL, add it as a Nexent custom MCP service, enable the six tools, debug the agent, and verify a post-publication question.

Technical Reference

MCP Server Configuration

{
  "mcpServers": {
    "crc-lnm-research-assistant": {
      "command": "uvx",
      "args": [
        "--index",
        "https://download.pytorch.org/whl/cpu",
        "crc-lnm-medical-agent-twomeme@1.1.1",
        "--transport",
        "stdio"
      ],
      "env": {}
    }
  }
}

Why the explicit CPU index?

It directs uvx to the CPU PyTorch wheels while keeping PyPI as the default index for the published MCP package and its other dependencies.

Why STDIO?

The published wheel contains the immutable model bundle and trusted release JSONL. initialize and list_tools construct only the lightweight MCP runtime. The model is loaded and integrity-checked on the first model operation; the case cache is built on the first case operation. Both use a platform temporary directory, so no repository, developer-machine path, or caller-supplied file is required.

All six tool schemas reject extra fields, including paths and URLs. Internal failures return a safe incident_id and failure_stage; these can be reported even where the platform does not expose server logs, without exposing stack traces or local paths.

docs/PLATFORM_DEPLOYMENT.md covers the separate authenticated Streamable HTTP container path. 使用说明.md documents the local release workflow and constraints.