cwtwb

Tableau workbook engineering for reproducible .twb / .twbx generation, validation, and migration.

cwtwb is a Python toolkit and Model Context Protocol (MCP) server for building Tableau Desktop workbooks from code or agent tool calls.

It is meant to be a workbook engineering layer, not a conversational analytics agent. The focus is reproducibility, inspectability, and safe automation in local workflows, scripts, and CI.

The cw in cwtwb comes from Cooper Wenhua.

Author: Cooper Wenhua <imgwho@gmail.com>

Website · Source · Changelog

Star History

Try the example workflow · Read the guide

Related MCP server: Tableau MCP Server

Quick Start

Install

pip install cwtwb

If you want the bundled Hyper-backed example too:

pip install "cwtwb[examples]"

If you want cloud validation (upload to Tableau Cloud/Server):

pip install "cwtwb[validate]"

Run As An MCP Server

uvx cwtwb

The short form above remains the simplest option and is the default config shown in this repository. cwtwb is a smart entrypoint: with no arguments in an interactive terminal it prints CLI help; when launched by an MCP client over stdio it starts the server.

Add the server to your MCP client with the same command. For example:

{
  "mcpServers": {
    "cwtwb": {
      "command": "uvx",
      "args": ["cwtwb"]
    }
  }
}

For Claude Code:

claude mcp add cwtwb -- uvx cwtwb

For VSCode, add cwtwb to your workspace or user mcp.json and use uvx cwtwb as the command.

If you prefer an explicit script name, these equivalent launch styles also work:

uvx cwtwb mcp
uvx --from cwtwb cwtwb-mcp
python -m cwtwb.mcp_server

Use As A CLI

The same package also exposes first-class command-line workflows for humans, scripts, CI, and agents that need direct file operations instead of MCP tool calls.

cwtwb --help
cwtwb doctor
cwtwb status --json
cwtwb inspect workbook.twb --json
cwtwb validate workbook.twb
cwtwb analyze workbook.twb --json
cwtwb run examples/specs/basic_cli.yaml

Common write commands require an explicit output path by default:

cwtwb create --out output/base.twb
cwtwb chart add output/base.twb --worksheet "Sales by Category" --mark Bar --rows Category --columns "SUM(Sales)" --out output/chart.twb
cwtwb dashboard add output/chart.twb --name Overview --worksheets "Sales by Category" --out output/dashboard.twb

Use --in-place only when you intentionally want to overwrite the input workbook, and --force only when replacing an existing output file.

MCP Client Stability

When cwtwb is connected as an MCP server, agents should call the exposed MCP tools directly through their client. They should not run shell commands such as mcp call cwtwb ..., mcp list-tools cwtwb, or gh api .../mcp/...; those commands are not part of cwtwb and are usually unavailable in normal Claude, Codex, Cursor, or VSCode environments.

If an agent cannot see tools such as create_workbook, add_worksheet, or save_workbook, restart or reconnect the MCP client and verify the server config. Clearing the uv cache only refreshes installed packages; it does not fix a stale client tool surface.

When using an existing .twb as a visual reference, agents must not copy Tableau XML column-instance tokens into chart inputs. Values such as [sum:Sales:qk], [none:Category:nk], [mn:Order Date:ok], or [federated.xxx].[sum:Profit:qk] are generated internals. Pass user-facing expressions such as Sales, SUM(Sales), Category, or MONTH(Order Date) instead.

Useful resources for agents:

cwtwb://tool-surface
cwtwb://skills/index
cwtwb://skills/dashboard_designer
file://docs/tableau_all_functions.json

Compatibility aliases are also available for common guessed URIs such as cwtwb://docs/manual-editing, but new prompts should prefer cwtwb://tool-surface and cwtwb://skills/index.

For client-specific details and the full reference, see https://github.com/aidatacooper/cwtwb/blob/main/docs/guide.md.

Dashboard Layout Files

Custom dashboard layouts can now be authored as either JSON or YAML using the same declarative DSL. For agent workflows, generate a layout file first, then pass that file path into add_dashboard(layout=...).

generate_layout_json("output/layout.json", layout_tree, ascii_preview)
generate_layout_yaml("output/layout.yaml", layout_tree, ascii_preview)

Both formats support the same wrapper structure:

• layout_schema: canonical dashboard layout tree

• _ascii_layout_preview: optional human/agent review aid

Highlights

See It In Action

This GIF shows the MCP tool flow that builds a dashboard step by step.

Architecture

                            Interfaces
  ┌───────────────────────────────────────────────────────────────┐
  │  ┌──────────────────────────┐  ┌───────────────────────────┐  │
  │  │        MCP Server        │  │      Python Library       │  │
  │  │  tools_workbook          │  │  from cwtwb.twb_editor    │  │
  │  │  tools_validate          │  │  import TWBEditor         │  │
  │  │                          │  │                           │  │
  │  │                          │  │  editor.add_...()         │  │
  │  │                          │  │  editor.configure_...()   │  │
  │  │                          │  │  editor.validate_schema() │  │
  │  │  (Claude / Cursor /      │  │  editor.save(...)         │  │
  │  │   VSCode / Claude Code)  │  │                           │  │
  │  └─────────────┬────────────┘  └──────────────┬────────────┘  │
  │                └──────────────┬────────────────┘               │
  └─────────────────────────────  ┼  ─────────────────────────────┘
                                  ▼
  ┌───────────────────────────────────────────────────────────────┐
  │                          TWBEditor                            │
  │       ParametersMixin  ·  ConnectionsMixin                    │
  │       ChartsMixin      ·  DashboardsMixin                     │
  │       validate_schema()  ·  save()                            │
  └──────────┬──────────────────┬──────────────────┬─────────────┘
             ▼                  ▼                  ▼
  ┌──────────────────┐  ┌──────────────┐  ┌──────────────────────┐
  │  Chart Builders  │  │  Dashboard   │  │  Analysis &          │
  │                  │  │  System      │  │  Migration           │
  │  Basic  DualAxis │  │              │  │                      │
  │  Pie    Text     │  │  layouts     │  │  migration.py        │
  │  Map    Recipes  │  │  actions     │  │  twb_analyzer.py     │
  │                  │  │  dependencies│  │  capability_registry │
  └────────┬─────────┘  └──────┬───────┘  └──────────┬───────────┘
           └───────────────────┼──────────────────────┘
                               ▼
  ┌───────────────────────────────────────────────────────────────┐
  │                    Packaged References                        │
  │    empty_template.twb  ·  Superstore XLS/Hyper                │
  │    tableau_all_functions.json  ·  dataset profiles            │
  │    vendored Tableau TWB XSD schemas (2026.1 / 2026.2)         │
  └───────────────────────────────┬───────────────────────────────┘
                                  ▼
  ┌───────────────────────────────────────────────────────────────┐
  │                     XML Engine  (lxml)                        │
  │    template.twb/.twbx  →  patch  →  validate  →  save        │
  └───────────────────────────────┬───────────────────────────────┘
                                  ▼
                      output.twb  /  output.twbx
                                  ▼
  ┌───────────────────────────────────────────────────────────────┐
  │               Cloud Validation (optional)                    │
  │    validate_workbook_api → REST API semantic validation      │
  │    upload_workbook       → Tableau Cloud/Server publish      │
  │    screenshot_workbook   → capture view for visual check     │
  └───────────────────────────────────────────────────────────────┘

Mermaid view:

flowchart TD
    subgraph Interfaces
        MCP["MCP Server<br/>tools_workbook<br/>tools_validate"]
        PY["Python Library<br/>TWBEditor API"]
    end

    subgraph Editor["Core Editor"]
        TWB["TWBEditor<br/>parameters · connections<br/>charts · dashboards<br/>validate_schema · save"]
    end

    subgraph Builders["Workbook Systems"]
        CHARTS["Chart Builders<br/>basic · dual-axis<br/>pie · text · map · recipes"]
        DASH["Dashboard System<br/>layouts · actions<br/>dependencies"]
        ANALYSIS["Analysis & Migration<br/>migration.py<br/>twb_analyzer.py<br/>capability_registry"]
    end

    subgraph References["Packaged References"]
        REFS["empty_template.twb<br/>Superstore XLS/Hyper<br/>Tableau functions<br/>TWB XSD schemas"]
    end

    subgraph Engine["XML Engine"]
        XML["lxml patch pipeline<br/>template.twb/.twbx → patch → validate → save"]
    end

    subgraph Outputs
        OUT["output.twb / output.twbx"]
        CLOUD["Cloud Validation<br/>REST semantic validation<br/>upload · screenshot"]
    end

    MCP --> TWB
    PY --> TWB
    TWB --> CHARTS
    TWB --> DASH
    TWB --> ANALYSIS
    CHARTS --> XML
    DASH --> XML
    ANALYSIS --> XML
    REFS --> TWB
    REFS --> XML
    XML --> OUT
    OUT --> CLOUD

The reference layer is packaged with the library so agents and scripts can start from known-good workbook assets, resolve Tableau calculation syntax, run Hyper-backed examples, and validate against local XSD schemas without relying on a checked-out repository.

Agent Architecture

cwtwb is designed for tool-using agents, not just direct Python calls. The MCP server gives agents a small, stateful workbook editing surface; skill resources give phase-specific Tableau guidance before each set of tool calls.

Human or agent prompt
        |
        v
MCP server instructions
        |
        v
Skill resources
calculation_builder -> chart_builder -> dashboard_designer -> formatting -> validation
        |
        v
Workbook tools
create/open -> list_fields -> add/configure -> layout -> save -> validate/upload
        |
        v
TWB/TWBX artifact + validation evidence

Prompts explain what to build. Skills explain how to build it well. Tools make the workbook changes inspectable and repeatable.

Capability Boundary

cwtwb keeps its public surface intentionally small:

Use list_capabilities or describe_capability when an agent needs to check whether a requested chart or workbook feature belongs in the stable surface.

Design Decisions

• The MCP server uses a stateful session model: open or create a workbook, mutate it through explicit tools, then call save_workbook.

• Skills are phase-specific operating guides, not generic prompt stuffing.

• save_workbook, validate_workbook, validate_workbook_api, and upload_workbook have separate responsibilities so agents do not confuse writing, local checks, semantic validation, and publishing.

• The capability registry keeps the product boundary explicit instead of letting showcase examples become accidental API promises.

Validation

cwtwb provides four levels of workbook validation:

# Level 1 — Local XSD (in-memory, no save required)
result = editor.validate_schema()
print(result.to_text())

# Level 3 — REST API semantic validation
from cwtwb.validate.uploader import TableauUploader
uploader = TableauUploader(env_path="project/.env")
result = uploader.validate("output.twb", validation_level="semantic")

# Save with local XSD validation; REST API semantic validation also runs when .env is configured
editor.save("output.twb")

# MCP tools
validate_workbook(file_path="output.twb")                                       # Local XSD validation
validate_workbook_api(twb_path="output.twb", validation_level="semantic")        # Default cloud semantic validation, no publish
validate_workbook_api(twb_path="output.twb", env_path="project/.env")            # Runtime credentials
upload_workbook(twb_path="output.twb")                                          # Publish/openability evidence or TWBX validation
screenshot_workbook(workbook_id="...", view_name="Sheet 1")                     # Visual check after upload_workbook

FAQ

What is the difference between .twb and .twbx?

.twb is the workbook XML. .twbx is the packaged version that bundles the workbook together with extracts and images.

Does validate_workbook save files?

No. validate_workbook() performs local XSD validation on the active in-memory workbook or an existing .twb / .twbx file. It does not write output. save_workbook() is the tool that writes files.

What validation does save() perform?

save() runs local XSD validation automatically before replacing the final output file. For .twb output, REST API semantic validation also runs when Tableau credentials are configured and the server supports it. Use validate_workbook_api(..., validation_level="semantic") when you want to request the Tableau Cloud/Server validation step directly.

What is upload_workbook for?

upload_workbook publishes a .twb or .twbx to Tableau Cloud/Server. Use it when you explicitly need publish/openability evidence, a workbook ID for screenshots, or .twbx package validation. For the default .twb cloud semantic check, prefer validate_workbook_api because it does not publish or store the workbook. Requires pip install "cwtwb[validate]" and Tableau credentials from environment variables, an explicit env_path, TABLEAU_ENV_FILE, or a .env file next to the workbook.

How do I set up Tableau Cloud/Server validation?

1. Install: pip install "cwtwb[validate]"

2. Copy .env.example to .env

3. Fill in your Tableau Cloud/Server PAT credentials

4. Call save_workbook to write the .twb or .twbx

5. Call validate_workbook_api for the default REST API semantic validation, or upload_workbook only when you also want publish/openability evidence, screenshots, or .twbx validation

Credential lookup order is explicit env_path first, then environment variables, TABLEAU_ENV_FILE, the workbook sibling .env, the current working directory .env, the cwtwb project .env, and finally the user's home .env. Prefer env_path for one-off MCP calls instead of editing MCP server configuration and restarting the server.

If validation reports that tableauserverclient is missing, call get_mcp_status first. It reports the MCP process's Python executable, cwtwb version, and whether the Tableau client is importable without exposing credentials. An env_path change is runtime-scoped and does not require an MCP restart; installing dependencies into a different Python environment does not fix the running server, so install the validation extra into the interpreter reported by get_mcp_status and reconnect only when the runtime or tool schema changes.

When should I use uvx cwtwb versus python -m cwtwb.mcp_server?

Use uvx cwtwb for the normal MCP workflow. Use python -m cwtwb.mcp_server for local testing without uvx.

For backward compatibility, uvx --from cwtwb cwtwb-mcp, python -m cwtwb.server, and python -m cwtwb.mcp continue to work.

Where is the full guide?

See the online guide.

Documentation

• Guide

• Changelog