Tabular models as tools for coding agents

Coding agents read files and write code well, but they can't learn from a table on their own. tabctx-mcp gives them TabPFN and TabICL as tools, so an agent can answer "which customers will leave, and why?" from a CSV. One MCP server works with Claude Code, Codex and Gemini CLI. The same tools also run as a JSON command line, and as a pi extension for agents without MCP.

TABCTX_BASE_URL is any tabctx deployment: one you run yourself (see the Quickstart), or Vector's gateway with a vp_ key, as below.

pip install "tabctx[mcp]"
export TABCTX_BASE_URL=https://proxy.vectorinstitute.ai TABCTX_API_KEY=vp_...
claude mcp add tabctx -- tabctx-mcp          # Claude Code (passes its environment on)
codex mcp add tabctx -- tabctx-mcp           # Codex; then in ~/.codex/config.toml, under
                                             #   [mcp_servers.tabctx]:
                                             #   env_vars = ["TABCTX_BASE_URL", "TABCTX_API_KEY"]
gemini mcp add tabctx tabctx-mcp \
  -e 'TABCTX_BASE_URL=$TABCTX_BASE_URL' -e 'TABCTX_API_KEY=$TABCTX_API_KEY'   # Gemini CLI
tabctx-mcp pi-extension > ~/.pi/agent/extensions/tabctx.ts                     # pi (no MCP)

The single quotes matter. Gemini CLI expands $VARS when it starts the server, so the key isn't written into its settings file. Agents without MCP or an extension can run the same tools as shell commands that print JSON: tabctx-mcp fit history.csv --target Churn, tabctx-mcp predict churn current.csv, and so on (tabctx-mcp --help).

Tool	What it does
inspect_table	Columns, their kinds, missing values and likely ids. With a target, the task and class balance.
evaluate	Held-out accuracy of each model next to a gradient-boosting baseline.
fit_model	Fits on every labelled row in seconds and saves the model under a name.
predict	Writes the table with prediction and probability columns, and returns the top-ranked rows.
what_if	Predicts hand-made rows, or one real row with some columns changed.
explain	Permutation importance: which columns the model relies on.
list_models	The models fitted so far, with their targets and features.
delete_model	Deletes a model: its context on the server and its saved files, which include its training table.

Fitted models live in .tabctx-mcp/ (or $TABCTX_MCP_HOME), so a model fitted in one call is still there for the next. tabctx-mcp skill prints the guide for agents. The readmission demo in examples/agent_demo/ runs the same task with each agent and scores their predictions against what the patients actually did.