MCP Server¶

Use Seldon straight from your AI assistant. The Seldon MCP server exposes Neuralk’s tabular foundation model to any MCP client — Claude Desktop, Claude Code, Cursor and more — so you can describe, predict, classify and evaluate tabular data through natural language, without leaving your tool.

Seldon uses in-context learning: you provide labeled examples as context and it predicts on new data, with zero hyperparameter tuning. The server wraps the neuralk SDK and handles authentication, data loading, and metrics for you.

Installation¶

Run the server directly from its repository with uv — no manual clone required:

uvx --from git+https://github.com/Neuralk-AI/mcp-server seldon-mcp

Or install it into your environment with pip:

pip install git+https://github.com/Neuralk-AI/mcp-server

You will need a Neuralk API key. Run neuralk login to create one — see the Quickstart for details. Keys look like nk_live_xxxxxxxxxxxx.

Connect to an MCP client¶

Claude Desktop¶

Add the server to claude_desktop_config.json:

{
  "mcpServers": {
    "seldon": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/Neuralk-AI/mcp-server", "seldon-mcp"],
      "env": {
        "NEURALK_API_KEY": "nk_live_your_api_key_here"
      }
    }
  }
}

Claude Code¶

claude mcp add seldon --env NEURALK_API_KEY=nk_live_your_api_key_here -- \
    uvx --from git+https://github.com/Neuralk-AI/mcp-server seldon-mcp

Remote (SSE)¶

Run the server as a shared, network-accessible endpoint:

seldon-mcp --transport sse --port 8000

Clients then connect over HTTP and pass their own API key with the x-neuralk-api-key header, so each user is billed against their own account:

{
  "mcpServers": {
    "seldon": {
      "url": "http://your-server:8000/sse",
      "headers": {
        "x-neuralk-api-key": "nk_live_your_api_key_here"
      }
    }
  }
}

Available tools¶

Tool	Description
`describe_data`	Load a tabular file and return a statistical summary — shape, column types, null counts, numeric stats, and sample rows. No API key required.
`predict`	Run predictions with Seldon. Provide a context file with labeled examples and either a separate file to predict on or an automatic holdout split.
`evaluate`	Run predictions and compute performance metrics against ground truth — accuracy, F1, precision and recall for classification; MAE, RMSE and R² for regression.
`list_models`	List the available Seldon model variants.

The server also ships guided prompts (classify, regress, compare_models) that walk an assistant through a full describe → predict → evaluate workflow, and exposes seldon://models and seldon://config as resources.

Supported file formats¶

CSV (.csv), Excel (.xlsx, .xls), Parquet (.parquet), and tabular JSON (an array of objects).

Model variants¶

Model	Description
`seldon-flash`	Optimized for low latency.
`seldon-small`	Balanced speed and accuracy (default).
`seldon-large`	Maximum accuracy for complex tasks.

Getting the most out of Seldon¶

Note

Seldon is not a traditional model where more data is always better. It learns from the context examples you provide, much like few-shot prompting. The context should be relevant to what you are predicting: to forecast churn for a specific customer segment, provide examples from that segment or very similar ones rather than the entire dataset. Irrelevant context adds noise and hurts performance. When your assistant helps you build a prediction, guide it to select the subset of your data that best represents the target.

Configuration¶

The server reads its configuration from environment variables (or a .env file):

Variable	Description
`NEURALK_API_KEY`	Server-level API key. Optional if clients provide their own via the `x-neuralk-api-key` header.
`SELDON_DEFAULT_MODEL`	Default model variant. Defaults to `seldon-small`.
`SELDON_DATA_DIR`	Base directory for resolving relative file paths. Defaults to the current directory.
`NEURALK_HOST`	On-premise inference server URL. Defaults to the Neuralk Cloud API. See On-Premise Deployment.