TableIntelligence *

A deterministic, reproducible intelligence layer for single-table data.

tabint is a Python library of statistical and machine-learning operations for one table at a time. Each operation is a plain, directly-callable function with a structured, inspectable result — plus an MCP server that exposes the same deterministic functions to any MCP-capable agent (Claude Cowork, Codex, Cursor).

The design goal that sets it apart: the same question yields the same, correct answer every time, with the method it chose made explicit. Code-generation tools that write fresh pandas on every run can't promise that; this library is built so the computation is deterministic and the statistical method is selected by transparent rules, not improvised.

Status: working library + MCP server. Phases 0–8 of the roadmap are implemented and tested (225 tests passing); see the roadmap below.

Install

The package is on PyPI. Requires Python ≥ 3.10.

# MCP server — no install needed, runs isolated via uvx
uvx --from tabint tabint-mcp --help

# or install the CLI + MCP server into your environment
pip install tabint

Verify it landed:

tabint --help        # the CLI
tabint-mcp --help    # the MCP server (stdio transport)

The package is named tabint; tabint-mcp is the server command inside it, so uvx/pipx need --from tabint to find the package. No uvx? Install it with curl -LsSf https://astral.sh/uv/install.sh | sh, or use pipx run --from tabint tabint-mcp --help instead.

Intended usage

Single table — the flat convenience API:

from tabint import Session

s = Session.load("customers.csv")
s.profile()                                   # describe every column
s.analyze_association("city", "spending")     # picks the right test by dtype
model = s.train_classifier(target="churn")    # returns a TrainedModel
model.predict(new_row)                        # predict lives on the model

Multiple related tables — one workspace, foreign keys detected automatically:

s = Session.load(["orders.csv", "customers.csv", "products.csv"])

s.relationships()                             # infers the FK graph:
#   orders.customer_id → customers.customer_id (100%)
#   orders.product_id  → products.product_id  (100%)

enriched = s.join(["orders", "customers"])    # materializes a new joined table
enriched.analyze_association("order_total", "tier")   # analytics run on it

s.table("customers").cluster()                # per-table handle for any table

Every analytic operates on one table — either an uploaded table or one produced by join. Joins are the only cross-table operation; they collapse related tables into a single table the rest of the library can reason about.

Use it from an AI agent (CLI + MCP)

The core is exposed to any MCP-capable agent through a terminal CLI (tabint) and an MCP server (tabint-mcp), both driven by a persistent session key. First do the install above, then register the server with your agent.

Environment variables

All agents need these in the server's env. Set them once and reuse the block in every config below.

Variable	Required	Default	Purpose
TABINT_API_KEY	yes	—	Your ti_… key from https://shubhamrandive.com/dashboard/account. Absent → free role (all analytics still work; persisting reports to the dashboard needs an account, enforced server-side).
TABINT_CONTROL_PLANE_URL	no	https://shubhamrandive.com	Base URL of the control plane (reports, folders, key validation).
TABULAR_BASE	no	current dir	Where on-disk sessions are stored (<base>/.tableint/sessions/).

Claude Code / Claude Cowork / Claude Desktop

Register the server (Claude Code CLI):

claude mcp add tabint \
  --env TABINT_API_KEY=ti_your_key_here \
  --env TABINT_CONTROL_PLANE_URL=https://shubhamrandive.com \
  -- uvx --from tabint tabint-mcp

…or paste the JSON block into the MCP config (Cowork / Desktop):

{
  "mcpServers": {
    "tabint": {
      "command": "uvx",
      "args": ["--from", "tabint", "tabint-mcp"],
      "env": {
        "TABINT_API_KEY": "ti_your_key_here",
        "TABINT_CONTROL_PLANE_URL": "https://shubhamrandive.com"
      }
    }
  }
}

OpenAI Codex (CLI)

Add to ~/.codex/config.toml (Codex reads MCP servers from [mcp_servers.*]):

[mcp_servers.tabint]
command = "uvx"
args = ["--from", "tabint", "tabint-mcp"]
env = { TABINT_API_KEY = "ti_your_key_here", TABINT_CONTROL_PLANE_URL = "https://shubhamrandive.com" }

Cursor

Add to .cursor/mcp.json in your project (or Settings → MCP for global):

{
  "mcpServers": {
    "tabint": {
      "command": "uvx",
      "args": ["--from", "tabint", "tabint-mcp"],
      "env": {
        "TABINT_API_KEY": "ti_your_key_here",
        "TABINT_CONTROL_PLANE_URL": "https://shubhamrandive.com"
      }
    }
  }
}

Verify it works

After registering the server in any agent, ask it to call the account_status tool — it should return your role:

> call account_status
{"role": "pro", "pro_features_unlocked": true, ...}   # or {"role": "free", ...} if no key set

Or from the CLI directly:

tabint load orders.csv customers.csv   # -> {"session_key": "s_ab12", "tables": [...], "relationships": [...]}
tabint associate order_total tier --session s_ab12 --table orders

See docs/agent-integration.md for the full tool list and troubleshooting. This replaces the originally-planned bespoke agent harness: any MCP-capable agent orchestrates the same deterministic functions.

Documentation

• docs/vision.md — what this is and why it exists
• docs/architecture.md — the layered design and contracts
• docs/algorithms.md — the full algorithm taxonomy
• docs/roadmap.md — phased build plan
• docs/adding-an-algorithm.md — the recipe for each new function
• CONTRIBUTING.md

Algorithm roadmap

Tick a box when a function is implemented, tested, and documented. This list is the single source of truth for "what to build next" — pick an unchecked item, research it, implement it against an existing library, add it to the test harness, then check it here.

Phase 0 — Foundation (build first; everything depends on it)

• store — load a table, run SQL, write columns back (DuckDB)
• results.Result — the structured return contract
• validation.dtypes — column type classification (the routing input)
• validation.assumptions — normality / equal-variance / sample-size checks
• identity — operation identity + caching key
• Session — state holder that delegates to the analytics layer
• eval harness — fixture CSVs + known-correct answers

Phase 1 — Descriptive

• profile — per-column type, distribution, missingness, cardinality, range
• detect_outliers — IQR and z-score flags
• association_matrix — pairwise association with the right measure per dtype

Phase 2 — Association / hypothesis testing (flagship — build carefully)

• analyze_association — dtype-routed test selection + effect size

Phase 3 — Clustering

• cluster — scale, fit, pick k (silhouette), write labels back as a column
• profile_clusters — characterize each cluster in plain terms

Phase 4 — Supervised learning

• train_classifier — fast lane, single model, proper split (returns TrainedModel)
• train_regressor — fast lane, single model, proper split
• backend="tabicl" — opt-in TabICL v2 tabular foundation model (in-context learning, no per-task training; needs the tabicl extra). Default backend="gbt".
• TrainedModel.predict / .predict_proba — bundled preprocessing
• evaluate — full metric set + confusion matrix
• add_predictions — write a model's predictions back as a column
• slow lane: AutoGluon wrapper as a job (infra ready via jobs; wrapper not yet written)
• jobs — Job registry + background runner

Phase 5 — Model interpretation

• feature_importance — gain-based / permutation importance
• explain_prediction — per-row SHAP values

Phase 6 — Dimensionality reduction

• reduce_dimensions — PCA, UMAP/t-SNE (PCA + t-SNE native; UMAP optional)

Phase 7 — Time series (optional; only if tables have a time axis)

• decompose — trend / seasonality / residual
• forecast — ARIMA / Prophet (ARIMA via statsmodels; Prophet optional)
• detect_changepoints — where a series shifts (ruptures; insights extra)

Phase 8 — Insight-extraction primitives (the "so what / why / what to do" layer)

• explain_metric — ranked key drivers + segment rules (shallow sklearn tree)
• market_basket — association-rule / cross-sell mining (mlxtend; insights extra)
• causal_effect — backdoor effect estimate + refutation (DoWhy; insights extra)
• rfm — Recency/Frequency/Monetary quintile segmentation (pandas)
• retention_cohorts — monthly cohort retention matrix (pandas)
• compare_periods — before/after shift with significance + effect size (scipy)

Later — large-data + orchestration (beyond V0)

• large-data strategies (sampling, out-of-core, approximate methods)
• natural-language ask() agent over the deterministic core

Scope

Every analytic operates on a single table. Multiple related tables can be loaded into one workspace (a shared DuckDB database); foreign keys are detected automatically and a join collapses related tables into a single derived table that the analytics then treat like any other. Reshaping beyond FK joins (pivots, complex multi-way transforms) remains upstream of where these algorithms begin.

License

Apache License 2.0 — see LICENSE. The distributed package (library, CLI, and MCP server) is fully open source. Monetization lives entirely in the hosted platform (the Stripe connector, cloud reports, and the Pro role), not in the client software.