Tableau Public Authoring MCP

An MCP server that authors Tableau workbooks from a declarative spec, validates them by opening them in the installed Tableau Public application, and publishes through Tableau's own visible desktop workflow.

This is the counterpart to the existing reader MCP (tableau-public), which inspects already-published content. The reader cannot create or publish anything. This server cannot browse public profiles. Keep both installed.

What it produces

Every screenshot below is a workbook generated from a JSON spec by this server, then opened in Tableau Public 2026.2.0 and photographed. Nothing was authored by hand, and nothing was touched in Tableau before the screenshot was taken. The data is the synthetic 12-row fixture in tests/fixtures/data/.

A dashboard — two generated worksheets tiled in one container, fixed 1200×800 canvas, title rendered:

A symbol map — latitude/longitude detected from the column names and tagged with Tableau's [Geographical] semantic role, which is what puts a real background map behind the marks. One mark per site, coloured by a dimension and sized by a measure:

A bar chart — colour by dimension, mark labels drawn, canvas fit applied:

The spec behind the map is the whole input:

{
  "name": "Site Map",
  "template": "map",
  "data_source": "sites",
  "detail": ["Site Name"],
  "color": "Status",
  "size": { "field": "Annual Sessions", "aggregation": "Sum" },
  "title": "Candidate sites"
}

Related MCP server: Tableau MCP Server

What it does

Verified capabilities

Everything below was confirmed by generating a workbook and opening it in Tableau Public 2026.2.0 on macOS, not merely by producing XML.

Supported

• CSV data sources, automatically materialised into .hyper extracts

• Field type inference: string, integer, real, boolean, date, datetime

• Worksheet templates: bar, line, text_table, heatmap, scatter, circle, map (all seven screenshot-verified in Tableau Public 2026.2.0, and re-checked on every live run by tests/live/test_render_templates.py)

• Geographic roles: latitude/longitude columns detected by name, which is what puts a real background map behind the map template

• Dashboards: title, fixed or automatic canvas, worksheets tiled in one container

• Calculated fields (Tableau formula syntax)

• Categorical filters, rendered as interactive filter cards

• Encodings: color, size, label, tooltip, and detail — detail raises the level of detail, so a scatter draws one mark per site instead of one per colour value

• Canvas fit: standard, fit_width, fit_height, entire_view

• Worksheet titles and subtitles

• .twbx packaging, with a byte-deterministic .twb inside

Not supported in v0.1 — see docs/architecture/CAPABILITY_BOUNDARY.md

• Stories, parameters, LOD expressions, joins/custom SQL, live connections

• Sets, groups, hierarchies, table calculations, reference lines

• Filled/choropleth maps, custom territories, map layers, spatial files

• Dashboard filter cards, legends, actions, floating zones, device layouts

Two claims to read carefully

• The .twbx is not byte-reproducible. The .twb inside it is — identical spec, identical bytes. The archive is not, because the embedded .hyper is written by Tableau's Hyper engine and carries non-deterministic internal state. Diff the .twb, never the package digest. See docs/LIMITATIONS.md §1.

• A standalone .twb is only usable next to its data directory. output.package_twbx: false writes <name>.twb plus a <name>.data/ directory holding each source's .hyper and CSV, and the .twb resolves its data relative to itself — so keep the two together, or move them together. A packaged build produces the same layout under Data/, matching what is inside the archive. Paths are always relative by necessity, not by preference: an absolute path would embed a home directory, and build_workbook refuses to emit machine-identifying data (see docs/architecture/THREAT_MODEL.md, T6).

Two constraints that shape everything

1. Tableau Public has no publishing API. The REST API covers Server/Cloud only, so publishing is driven through the real UI. Tableau documents the command as Server > Tableau Public > Save to Tableau Public, but that is the Desktop Professional path: in the Public edition (2026.2.0) the Server menu holds only Run Optimizer... and the command lives on the File menu. It is invoked with a real Command+Shift+S key event, with the File menu as a fallback — see docs/PUBLISHING_WORKFLOW.md.

2. Tableau Public rejects non-extract data sources (Error Code: 3C242D89). Every CSV is therefore converted to a Hyper extract at build time.

Install

uv tool install git+https://github.com/andyyaro/tableau-public-authoring-mcp
tableau-public-authoring doctor

No clone, no virtualenv, no paths to edit. doctor checks everything the server needs and prints exactly what to fix — run it before registering, because every prerequisite below fails silently at runtime rather than raising:

  [PASS] platform          macOS
  [PASS] python            3.13.14 at ~/.local/share/uv/tools/…/bin/python
  [PASS] mcp sdk           importable
  [PASS] tableauhyperapi   importable
  [PASS] quartz (pyobjc)   importable
  [PASS] tableau public    2026.2.0 at /Applications/Tableau Public.app
  [PASS] accessibility     granted to this process
  [PASS] screen recording  capture succeeded

macOS permissions are granted per executable, so grant them to the interpreter doctor prints — not to Tableau, and not to your terminal if you register a different one. Details in docs/MACOS_PERMISSIONS.md.

Requirements: macOS with Tableau Public installed (developed against 2026.2.0, Apple silicon), Python ≥ 3.10, and uv.

Register with Claude Code

claude mcp add --transport stdio --scope user tableau-public-authoring \
  -- tableau-public-authoring

That is the whole command — uv tool install puts the executable on your PATH at a stable location, so there is no absolute path to substitute and it survives upgrades. Registering does not affect any other MCP server; verify with claude mcp list. Restart Claude Code before the tools appear.

git clone https://github.com/andyyaro/tableau-public-authoring-mcp
cd tableau-public-authoring-mcp
uv venv && uv sync && uv pip install -e .
uv run tableau-public-authoring doctor

Register that build by pointing at its interpreter explicitly:

claude mcp add --transport stdio --scope user tableau-public-authoring \
  -- /absolute/path/to/.venv/bin/python -m tableau_public_authoring.server

Avoid uvx --from … for the registered server: it runs from a content-hashed cache path that changes whenever a dependency changes, and the macOS permission grants are tied to that path — so the server would quietly lose Accessibility and Screen Recording on the next upgrade. It is fine for one-off use such as uvx --from … tableau-public-authoring doctor.

Upgrading

uv tool upgrade tableau-public-authoring
tableau-public-authoring doctor

Build a workbook

{
  "spec_version": "0.1",
  "workbook": { "name": "EV Charging Readiness", "description": "Generated" },
  "data_sources": [{ "id": "sites", "type": "csv", "path": "/abs/path/sites.csv" }],
  "calculated_fields": [{
    "name": "Readiness Band", "data_source": "sites", "datatype": "string",
    "formula": "IF [Readiness Score] >= 70 THEN \"High\" ELSE \"Low\" END"
  }],
  "worksheets": [{
    "name": "Readiness by Site", "template": "bar", "data_source": "sites",
    "rows": ["Site Name"],
    "columns": [{ "field": "Readiness Score", "aggregation": "Sum" }],
    "color": "Readiness Band",
    "filters": ["Status"],
    "title": "EV Charging Readiness by Site"
  }],
  "output": { "directory": "/abs/path/output", "package_twbx": true }
}

Paths may be relative to a single base_dir instead of absolute — and a relative path is confined to that directory, which an absolute path is not. See docs/AUTHORING_SPEC.md.

Then: validate_workbook_spec (dry run) → build_workbook → validate_in_tableau_public → publish_to_tableau_public.

Full schema: schemas/workbook-spec-v0.1.json and docs/AUTHORING_SPEC.md.

Why things are the way they are

Most of this codebase's shape is a response to something Tableau actually did. docs/LIMITATIONS.md lists the constraints; docs/history/CHANGE_HISTORY.md keeps the release write-ups, which record the failure each fix was responding to and what proved it fixed. docs/fixtures/ holds the forensic analysis of a real published workbook that the compiler was reverse-engineered from.

Publishing

publish_to_tableau_public is marked destructive and requires both confirm_publish: true and profile (your Tableau Public profile name); overwrite defaults to false. profile is required because without it the server cannot check whether a workbook of that name is already live, and Tableau Public replaces a same-named workbook irreversibly, with no version history.

It drives the visible command and never reads, types, or stores credentials — it uses your existing signed-in desktop session. If Tableau needs a sign-in, the tool stops and tells you to sign in manually. Publication is irreversible and world-visible.

Use dry_run: true to see exactly what would happen first.

Tests

uv run pytest                    # unit + integration (no Tableau needed)
uv run pytest -m live            # requires Tableau Public installed
uv run ruff check . && uv run ruff format --check .

live and publish tests are excluded from normal runs and must never run in CI.

Security

No credentials are stored anywhere. XML is generated through ElementTree (never string concatenation), archives are ZIP-slip guarded, paths are traversal-checked, and generated workbooks are scanned for machine-identifying data before packaging. See docs/SECURITY.md and docs/architecture/THREAT_MODEL.md.

Rollback

claude mcp remove --scope user tableau-public-authoring

This leaves all other MCP servers untouched.

Uninstalling

claude mcp remove tableau-public-authoring -s user
claude mcp list          # your other servers should still be listed
rm -rf <this repo>/output   # generated workbooks are inert local files

Anything already published must be deleted by you from your Tableau Public profile in the browser. This server has no delete capability, by design: publication is irreversible from its side, which is why it requires confirm_publish and refuses a name collision unless you pass overwrite.

License

MIT — see LICENSE.