TabPFN MCP, gives LLMs tools for predictions on tabular data

Exposes TabPFN's pre-trained neural network for tabular prediction as an MCP tool that LLMs can invoke directly. The capability works by wrapping TabPFN's inference engine as a callable MCP resource, allowing language models to request predictions on structured data without leaving the conversation context. Requests flow through the MCP protocol layer, which serializes tabular inputs (features, targets, test samples) and routes them to TabPFN's transformer-based backbone for few-shot learning on new datasets.

Leverages TabPFN's pre-trained transformer architecture to make accurate predictions on new tabular datasets with minimal examples (few-shot learning), without requiring model fine-tuning or retraining. The capability works by using TabPFN's learned inductive biases for tabular data to generalize from small labeled datasets. When invoked via MCP, the tool accepts a small training set and immediately produces predictions on test samples, with the model's weights remaining frozen.

Registers TabPFN's prediction capability as a callable MCP tool with a standardized schema, enabling any MCP-compatible LLM client to discover and invoke tabular predictions. The tool registration includes a JSON schema describing input parameters (training data, test data, task type), output format, and usage instructions. When an LLM decides to use the tool, the MCP protocol serializes the request, routes it to the TabPFN backend, and deserializes results back into the conversation context.

Automatically detects whether a tabular prediction task is classification or regression based on the target variable's properties (discrete vs. continuous), and routes the request to the appropriate TabPFN inference path. The capability examines the training target's data type and cardinality to determine task type, then configures the model's output layer and loss function accordingly. This abstraction allows LLMs to specify prediction tasks without explicitly declaring the task type.

Maintains prediction context across multiple MCP tool invocations within a single LLM conversation, allowing the agent to reference previous predictions, refine datasets, or chain multiple prediction tasks. The capability preserves training data, previous predictions, and metadata across tool calls, enabling the LLM to build on prior results without re-specifying the full dataset each time. Context is managed at the MCP protocol level, with each invocation including references to prior state.

Provides prediction uncertainty estimates alongside point predictions, leveraging TabPFN's learned ability to quantify confidence on new tasks. The capability returns confidence scores, prediction intervals, or probability distributions for each prediction, derived from the model's internal representations and learned uncertainty patterns. This allows LLM agents to assess prediction reliability and make risk-aware decisions based on confidence thresholds.

Enables LLM agents to compare predictions across different feature subsets, training set sizes, or dataset variations within a single workflow, supporting ablation studies and feature importance analysis. The capability allows the agent to invoke TabPFN multiple times with different inputs (e.g., full feature set vs. reduced feature set) and aggregate results for comparison. Results are structured to highlight differences and enable the agent to reason about feature contributions or data quality impacts.