MCP server¶

coreai-onnx ships an MCP server so agent frameworks can call the converter as native tools.

Install and configure¶

pip install "coreai-onnx[mcp]"

Add to your MCP client configuration (Claude Code, Claude Desktop, etc.):

{
  "mcpServers": {
    "coreai-onnx": {
      "command": "coreai-onnx-mcp"
    }
  }
}

The server speaks stdio. It exposes four tools:

Tool	Parameters	Wraps
`inspect_model`	`model_path`	`coreai-onnx inspect`
`convert_model`	`model_path`, `output_path`, `optimize=true`, `validate=true`, `verify=true`, `rtol`, `atol`, `min_psnr`, `compute_unit`, `seed`, `entrypoint`	`coreai-onnx convert`
`verify_model`	`model_path`, `aimodel_path`, `rtol`, `atol`, `min_psnr`, `compute_unit`, `seed`, `entrypoint`	`coreai-onnx verify`
`get_schema`	—	`coreai-onnx schema`

One contract¶

Every tool returns the same envelope the CLI emits with --json (schema_version, command, status, result, warnings, error) — see Machine-readable output. Domain failures (unsupported ops, bad files, precision failures) come back as status: "error" envelopes with the documented error codes, never as MCP protocol errors; branch on error.code exactly as you would for the CLI. Exit codes do not exist over MCP — status/error.code carry the same information. Boolean parameters map to the CLI’s negative flags (optimize=false ≙ --no-optimize); entrypoint ≙ --name.