crosstabs.AI workspace
Connect an agent

Agent-native research production

The research workspace, now callable.

Give an agent the whole research job—not a chat box beside a table. Import, profile, define, test, compare, code, review, refresh, and deliver with the evidence and audit trail intact.

Connect an agent Open browser workspace

23 lifecycle tools · Local full lifecycle · Hosted live-readiness gated

Live product demonstration · revision 12

One agent run, inside the real workspace model.

Switch modes, inspect the evidence, and ask a follow-up. The demonstration uses the same workspace controls and deterministic result components.

crosstabs.

Brand tracker · Q3

Local MCP · on-device · revision 12

Audited

Analysis canvas

Advocacy × Audience

CountColumn %

Applied research defaults

  • Base

    Weighted n = 2,408

    Weighted; verify design effect

  • Missing

    Excluded

    Row and column missing values

  • Weight

    Survey weight

    Approximate inferential tests

  • Tests

    Automatic

    χ² + corrected column comparisons

Column % · weighted · letters at 95%

N = 2,408 · Effective N = 2,181

Cramér's V

0.351

Cross-tabulation of Advocacy (rows) by Audience (columns). Column headers are sortable.
Advocacy(A)(B)Total
Recommend1,264
Do not recommend1,144
Total1,2041,2042,408

Column tests: letters mark columns each cell's column % is significantly higher than (uppercase 95%, lowercase 90%, Bonferroni-corrected pairwise z-tests; columns with base < 30 excluded).

Adjusted residual highlights: green is above expected and red is below expected at |residual| > 1.96.

Evidence-linked finding: Customers are significantly more likely to recommend than prospects. Association, not causation.

χ²(1) = 296.4p < .001audit event 18export ready

One contract for the browser workspace, local agents, and explicitly authorized hosted automation.

The workspace is the contract

One research loop. Every surface.

The browser and agent share the same project stages, definitions, calculations, approvals, revisions, and editable outputs—not merely the same tool names.

  1. 01

    Project

  2. 02

    Data

  3. 03

    Analyze

  4. 04

    Findings

  5. 05

    Outcomes

  6. 06

    AI

  7. 07

    Report

Three ways to run the same research job.

Local privacy · Hosted persistence · Governed generation

01

On-device · full lifecycle

Local MCP

Import local files, run row-level projects, tab books, waves, coding review, editable reports, export, and audit history without uploading the source dataset.

02

Live inventory · fail closed

Hosted MCP

The endpoint advertises protected project tools only while its database, OAuth metadata, scopes, and signing keys pass live readiness. Otherwise it exposes only five public evidence and aggregate tools.

03

Generative work only

Vercel AI Gateway

Open-end theme proposals use a disclosed live model with evidence, approval, and undo. Deterministic statistics never depend on an LLM.

Choose the boundary

Private locally. Persistent when hosted.

Use local stdio when files must stay on-device. Before a hosted project job, inspect the live inventory: project tools appear only after the full production readiness gate passes.

View verified package on PyPI

Local · confidential projects

Python 3.10+ · Node 20+
pip install crosstabs
crosstabs-headless

Hosted · live inventory

Streamable HTTP · tools/list is authoritative
https://mcp.crosstabs.com/mcp

Live MCP inventory is authoritative

Checking the production readiness gate…

Try a complete job

Agent prompt
Create a project from survey.csv. Profile cardinality, define age and region banners, run a weighted tab book, code verbatims for approval, and export an editable PPTX with audit evidence.

Progressive reference

23 tools. One versioned contract.

Tool count is not parity evidence. Calculations, defaults, state, warnings, artifacts, and audit behavior are release-gated through cross-surface fixtures.

01

Project

5 tools
  • create_project

    Create an empty, versioned research project.

  • import_dataset

    Validate and import respondent records into a project wave.

  • profile_dataset

    Return bounded types, missingness and cardinality diagnostics.

  • list_projects

    List redacted project summaries available to the caller.

  • inspect_project

    Inspect a redacted project graph, capabilities and current revision.

02

Design

4 tools
  • define_row_set

    Create or replace a reusable ordered row-variable set.

  • define_banner

    Create or replace a reusable multilevel banner.

  • apply_filter

    Create and assign an auditable project filter.

  • set_weight

    Create and assign a validated weighting definition.

03

Analyze

3 tools
  • run_table

    Compute one crosstab with diagnostics and evidence metadata.

  • run_tab_book

    Compute the selected tab-book recipe with progress and cancellation.

  • compare_waves

    Compare compatible tables across project waves.

04

Code

4 tools
  • code_open_ends

    Use Vercel AI Gateway to propose evidence-linked themes and assignments.

  • review_themes

    Approve, reject or edit proposed coding themes.

  • approve_coding

    Apply an approved coding run as auditable project variables.

  • undo_change

    Reverse a recorded reversible project mutation.

05

Track

3 tools
  • replace_dataset

    Import a replacement wave without overwriting prior source data.

  • detect_schema_drift

    Inspect wave compatibility and propose repairs without mutation.

  • repair_schema

    Apply an approved, reversible schema repair plan.

06

Deliver

4 tools
  • generate_report_pack

    Generate an editable evidence-linked PPTX or DOCX pack.

  • refresh_report_pack

    Refresh generated regions while preserving declared manual regions.

  • export_project

    Export a definition-only package by default or explicitly authorized full data.

  • get_audit_history

    Return paginated append-only project audit events.

Trust is part of the result

Inspectable by design.

  • Append-only audit history across local and hosted project mutations.
  • Human approval before coding application, raw-row transfer, and high-impact changes.
  • Deterministic tables with bases, effect sizes, warnings, and evidence links.
  • Actor ownership, exact project-bound access, separate row-transfer grants, and transparent local mode.