Skip to main content
The TabPFN Prediction Skill teaches Claude how to run classification and regression predictions on your tabular data — CSVs, spreadsheets, or tables pasted directly into chat — without writing any ML code. Instead of manually calling MCP tools and formatting parameters, you describe what you want in natural language and Claude handles the rest. Claude Skills are markdown instruction files that extend what Claude can do. This skill gives Claude a structured playbook for working with the TabPFN MCP server: it knows which tools to call, validates your data before making requests, confirms task details with you, and handles edge cases like missing values or mismatched columns.

GitHub Repository

Get TabPFN Prediction Skill here

​
Prerequisites

Before installing the skill, make sure you have:
  • A Prior Labs account — sign up at ux.priorlabs.ai if you don’t have one
  • TabPFN MCP server configured in your Claude client — follow the MCP setup guide
  • Claude Code installed or Claude Desktop on a Pro, Max, Team, or Enterprise plan

​
How it works

When you ask Claude to predict outcomes on tabular data, the skill activates and guides Claude through a structured workflow:
  1. Confirms details with you — target column, task type (classification or regression), and train/test split. Claude won’t guess these.
  2. Validates your data — catches issues that would cause errors (missing targets, column mismatches, duplicate column names) while leaving things TabPFN handles natively alone (missing feature values, mixed types, unnormalized numerics).
  3. Calls the right MCP tools — fit_and_predict for inline tables pasted into chat, upload_dataset followed by fit_and_predict_from_dataset for CSV files, or predict to reuse a previously fitted model on new data.
  4. Returns results — predictions, class probabilities, or regression estimates, formatted for your context.
Without the skill, Claude can still use the TabPFN MCP tools directly, but it won’t know the optimal workflow, data validation rules, or how to handle edge cases like high-cardinality categoricals or class imbalance. The skill encodes this domain knowledge so every prediction follows best practices. The skill activates automatically when you use phrases like “predict X from this data”, “classify these rows”, “estimate”, “forecast”, or mention TabPFN. In Claude Code, you can also invoke it directly with /predicting-with-tabpfn.

​
Install the skill

  1. Download the SKILL.md file from the repository.
  2. Open the sidebar and click Customize.
Claude Desktop sidebar with Customize option
  1. Click Skills in the left panel.
Customize page showing Skills and Connectors
  1. Click the + button at the top of the skills list.
Skills list with the plus button
  1. Select Upload skill from the dropdown.
  2. Select the SKILL.md file you downloaded in step 1.
The skill should appear under My skills once uploaded.

​
Allow network access

The TabPFN MCP server runs remotely at api.priorlabs.ai. When Claude calls TabPFN tools, it needs permission to make outbound network requests to this domain.
When Claude Desktop runs code in its local coding environment, network access is restricted by default. To allow connections to the TabPFN API:
  1. Open Settings in the sidebar
  2. Enable Code execution if you haven’t already
  3. Under Network egress, toggle Allow network egress on
  4. Add api.priorlabs.ai to the allowed domains list
Claude Desktop will also prompt you to authorize MCP tool usage the first time. Accept the prompt to allow TabPFN predictions.
On Team and Enterprise plans, an admin may need to configure the network allowlist. Domains like github.com and pypi.org are allowed by default, but api.priorlabs.ai must be added explicitly.

​
Make your first prediction

Once the MCP server is configured and the skill is installed, start a conversation with Claude and try one of the examples below.

​
Inline data

Paste a small table directly into chat: 
Here's my dataset. Predict the "outcome" column.

| age | income | credit_score | outcome  |
|-----|--------|-------------|----------|
| 25  | 50000  | 680         | approved |
| 45  | 120000 | 750         | approved |
| 33  | 35000  | 580         | denied   |
| 52  | 90000  | 710         | approved |
| 28  | 42000  | 620         | denied   |
| 38  | 75000  | 695         |          |
| 47  | 105000 | 720         |          |
Claude will:
  1. Identify outcome as the target column
  2. Confirm this is a classification task
  3. Use rows with filled targets as training data, empty rows as test data
  4. Call the TabPFN MCP tools and return predictions for the missing rows

​
CSV file

Point Claude to files in your project, or drop them directly into the chat: 
I have train.csv and test.csv in the data/ folder.
Predict the "target" column using the training set.
Claude will upload both files to the TabPFN server, fit a model on the training data, and return predictions for the test set.

​
What you can do

CapabilityDetails
ClassificationBinary or multi-class with optional probability output
RegressionContinuous values with mean, median, mode, or quantile output
Inline dataPaste tables directly into chat for quick predictions
File-based dataCSV files for larger datasets
Re-predictionReuse a fitted model on new data without re-training