qlik-cloud-mcp

Qlik Cloud MCP — Capability Registry

What This Skill Covers (and What It Doesn't)

The MCP tool definitions (visible to every agent at invocation time) already document each tool's parameters, response structures, usage examples, and basic notes. This skill does NOT repeat that information.

This skill covers:

• MCP detection — how agents determine MCP availability and branch
• Pipeline phase mapping — which tools are relevant at which phase
• Behavioral gotchas — issues discovered through live testing that go beyond the tool definitions (see references/behavioral-notes.md for the full list)
• Workflow patterns — multi-step tool call sequences for common operations
• Framing — what the MCP server can and cannot do, to prevent incorrect assumptions

What the Qlik Cloud MCP Server IS and IS NOT

IS: The Qlik Cloud platform API exposed through MCP. It operates on data and objects already in Qlik Cloud: apps, datasets in the catalog, master items, sheets, glossaries, data products, and lineage metadata.

IS NOT: A direct connection to upstream source databases. It cannot run SQL queries against SQL Server, PostgreSQL, or any source system. Source database profiling (Phase 2, Path A in the source-profiler skill) requires separate database MCP connections. If those aren't available, the source-profiler's manual template path (Path B) applies.

The practical distinction: When the framework says "MCP available" for source profiling, it means a database MCP connection, not the Qlik Cloud MCP server. The Qlik Cloud MCP server enables a different set of capabilities: inspecting what's already loaded in Qlik apps, validating expressions against live data, scaffolding visualizations, profiling Qlik-managed datasets (QVDs, cloud data files), and tracing lineage.

MCP Detection

Before using any tool in this registry, agents must determine whether the Qlik Cloud MCP tools are available in the current session. Tools are prefixed with qlik_.

IF qlik_* tools are present:
    Qlik Cloud MCP is live → use MCP-enhanced workflows
ELSE:
    Qlik Cloud MCP unavailable → use manual fallback workflows
    Log "Qlik Cloud MCP unavailable" in .pipeline-state.json

MCP availability can change between sessions. Always check, never assume.

Tool-to-Pipeline-Phase Mapping

This table maps every MCP tool to the pipeline phases where it's useful. Tools not listed here are governance tools documented in Section 6.

Tool	Phase 0	Phase 2	Phase 3	Phase 4	Phase 5	Phase 6	Phase 7
describe_app	●
get_fields	●		●	●			●
get_field_values		●	●		●		●
search_field_values	●		●		●		●
list_sheets	●					●
get_sheet_details	●					●	●
list_dimensions	●				●	●
list_measures	●				●	●
create_data_object				●	●		●
get_chart_data					●		●
get_chart_info	●						●
select_values					●		●
clear_selections					●		●
get_current_selections					●		●
create_sheet						●
add_chart						●
add_filter						●
create_dimension						●
create_measure					●	●
search	●	●
get_dataset		●
get_dataset_schema		●
get_dataset_sample		●					●
get_dataset_profile		●					●
get_dataset_freshness		●					●
get_dataset_trust_score							●
get_lineage	●	●
search_spaces	●
search_users	●

Reading the table: ● means the tool has a defined use in that phase. Agents should consult the relevant workflow pattern (Section 5) for the sequencing.

Key Behavioral Notes

These are gotchas discovered through live testing that are NOT documented in the tool definitions. For the complete per-tool list, read references/behavioral-notes.md.

Silent failures are the biggest risk. Several tools return null/zero instead of errors when given invalid input:

• create_data_object: Referencing a non-existent field returns null/0, not an error. Always verify field names with get_fields first.
• select_values: Selecting a value that doesn't exist in the field produces an empty selection silently. The returned selection list is the ground truth -- if the field doesn't appear in it, the selection failed.
• create_measure / add_chart: Invalid expressions succeed at creation time but fail at render time. Always validate expressions with create_data_object before creating permanent objects.

Selection state is persistent and cumulative. select_values filters the entire session. Every subsequent create_data_object, get_field_values, and get_chart_data call reflects active selections. Always call clear_selections after validation runs. Some app-level default selections (script-embedded, Section Access, default bookmarks) cannot be cleared via MCP. Always call get_current_selections before profiling or validation work and document any active selections. When profiling, if selections cannot be cleared, consider using {1} set analysis to ensure data is profiled across all values, not only those currently selected.

Visualization creation has no layout control. create_sheet, add_chart, and add_filter create objects but provide zero control over grid positioning, sizing, visual formatting, colors, number formats, or conditional formatting. Objects are appended sequentially. The manual build checklist remains essential for final layout work.

Master items in published apps. Master dimensions and measures can only be created/edited on private sheets. In published apps (managed spaces), changes must be made in the source app and re-published. The MCP tools don't enforce this restriction -- they'll create items that may not be editable later. Additionally, if a master measure is renamed or deleted, any expressions referencing it by name will return NULL rather than raising an error.

Hypercube cell limit. create_data_object and get_chart_data are subject to a 10,000 cells per request limit (dimensions × measures × rows). For large result sets, use pagination (offset/limit) and keep dimension cardinality bounded with sort + limit settings.

Lineage is one level per call. get_lineage returns only immediate upstream dependencies. To trace a full pipeline (Source → Extract → QVD → Transform → App), call recursively on each upstream node. QRI format for apps: qri:app:sense://[appId]. Get dataset QRIs from get_dataset response.

Trust scores are often absent. get_dataset_trust_score returns an error (not null) when no trust score exists. Handle this gracefully -- most datasets won't have one unless explicitly assessed.

Workflow Patterns

These are the multi-step tool call sequences agents should follow. Each pattern lists the tools in order with the reasoning for each step.

5.1 Reference App Analysis (Phase 0)

Use when analyzing an existing app to extract patterns for replication.

1. describe_app(appId)
   → App overview, field count, metadata. First call for any app.

2. get_fields(appId)
   → Full field inventory. Note $key, $date, $numeric tags for
   type classification and association detection.

3. list_sheets(appId)
   → Sheet inventory. Get IDs for step 4.
   ⚠ Do NOT rely on the cells array here -- it's truncated.

4. For each sheet: get_sheet_details(appId, sheetId)
   → Complete object list per sheet. Extract chart types,
   dimension/measure assignments.

5. list_dimensions(appId) + list_measures(appId)
   → Master item inventory with definitions and expressions.

6. get_lineage(qri:app:sense://[appId])
   → Trace upstream pipeline. Recurse on each upstream node
   for full lineage chain.

5.2 Expression Validation (Phase 5)

Highest-value MCP integration. Validates expressions from the expression catalog against loaded data.

1. clear_selections(appId)
   → Ensure clean state. Verify with get_current_selections.

2. get_fields(appId)
   → Build field name lookup. Every field reference in every
   expression must match exactly (case-sensitive).

3. For each expression in the catalog:
   a. create_data_object with the expression and a relevant dimension
   b. Check result:
      - Non-null → expression evaluates ✓
      - Null/0 → could be valid (no matching data) or invalid
        (field typo, syntax error). Cross-check with
        get_field_values to confirm data exists.

4. For set analysis expressions, test with known-good values:
   - Sum({<Year={2024}>} [Amount]) with a year known to have data
   - Verify the filter produces expected narrowing

5. For comparative expressions (YoY, variance):
   a. select_values to set a known context
   b. Evaluate the expression
   c. clear_selections immediately after

6. Produce validation summary:
   | Expression | Expected | Actual | Status | Notes |
   |---|---|---|---|---|
   | vRevenue | Numeric > 0 | 1,234,567 | ✓ Pass | |
   | vMargin | 0-100% | null | ⚠ Investigate | Check field name |

5.3 Visualization Scaffolding (Phase 6)

Creates the skeleton of sheets and objects. Manual work is still required for layout and formatting.

1. Create master dimensions:
   For each dimension in master item definitions:
     create_dimension(appId, name, dim, description)
     → Record returned libraryId

2. Create master measures:
   For each measure in expression catalog:
     create_measure(appId, name, expr, description)
     → Record returned libraryId
   ⚠ Validate expressions with create_data_object BEFORE this step.
   Invalid expressions succeed here but fail at render.

3. Create sheets:
   For each sheet in viz specifications:
     create_sheet(appId, title, description)
     → Record returned sheetId

4. Add filter panes, then charts:
   For each sheet:
     add_filter(appId, sheetId, title, fields)
     add_chart(appId, sheetId, chartType, title,
       dimensions=[{libraryId: dimId}],
       measures=[{libraryId: measureId}])

5. Document what requires manual completion:
   - Grid positioning and object sizing
   - Visual formatting (colors, fonts, number formats)
   - Conditional formatting and calculation conditions
   - Responsive behavior adjustments

5.4 Data Quality Validation (Phase 7)

Validates data quality through MCP as part of comprehensive QA.

1. clear_selections(appId)

2. Null rate checks per key field:
   create_data_object(measures=[
     {expression: "Count([Field])", label: "Non-Null"},
     {expression: "NullCount([Field])", label: "Nulls"}
   ], dimensions=[])

3. Uniqueness checks on key fields:
   create_data_object(measures=[
     {expression: "Count([KeyField])", label: "Total"},
     {expression: "Count(DISTINCT [KeyField])", label: "Distinct"}
   ], dimensions=[])
   → Total != Distinct means duplicates. Critical finding.

4. Encoded null scan:
   search_field_values(searchTerms=["N/A", "NULL", "TBD",
     "-", "Unknown", "NONE", "n/a"])
   → Searches all fields for common null encodings.

5. For Qlik-managed datasets, augment with catalog tools:
   get_dataset_profile(datasetId) → pre-computed statistics
   get_dataset_freshness(datasetId) → data currency
   get_dataset_trust_score(datasetId) → may return error (handle gracefully)

5.5 Post-Reload Spot Checks (Phase 4)

Quick validation that data loaded correctly after a Qlik reload.

1. get_fields(appId)
   → Verify expected fields exist. Check $numeric/$text/$date tags
   match expected types.

2. Row count checks:
   create_data_object(measures=[
     {expression: "Count([Table.KeyField])", label: "Rows"}
   ], dimensions=[])
   → Compare against expected row counts from source profiling.

3. Spot-check field values:
   get_field_values(appId, fieldName, limit=20)
   → Verify sample values look reasonable.

Governance Tools

The MCP server includes tools for Qlik Cloud governance features (business glossaries and data products). These don't have a mandatory pipeline phase but provide value for organizations using Qlik's governance capabilities.

Business Glossary tools (create_glossary, create_glossary_term, create_glossary_category, search_glossary_terms, get_glossary_term, update_glossary_term, update_term_status, delete_glossary_term, get_glossary_categories, get_glossary_term_links, create_glossary_term_links, get_full_glossary_export): Create and manage governed business term definitions. Terms can be linked to apps, datasets, fields, master dimensions, and master measures via create_glossary_term_links. Term status follows a draft → verified → deprecated lifecycle where only stewards can verify or modify verified terms. Linking glossary terms to master items brings business context into the app for end users. As a best practice, place glossaries in a shared space where all tenant users have Can view access.

Data Product tools (create_data_product, get_data_product, get_data_product_documentation, update_data_product, update_data_product_space, update_activate_data_product, update_deactivate_data_product, delete_data_product): Curated, governed bundles of datasets. Support a draft → active → deactivated lifecycle with key contacts, tags, and markdown documentation.

Dataset metadata tools (update_dataset_metadata, update_dataset_quality, get_dataset_quality_computation_status, get_dataset_memberships): Update dataset descriptions/tags, trigger quality computations, and check data product memberships.

Potential framework usage:

• Phase 8 (Documentation): Create glossary entries from the data dictionary, linking terms to master items created in Phase 6.
• Post-pipeline: Package pipeline outputs into governed data products. Update dataset metadata with framework-generated descriptions.

The tool definitions document these tools' parameters thoroughly. Consult them directly when using governance features.

Extensibility

When the Qlik Cloud MCP server adds new tools:

1. Test the tool against a live tenant to understand actual behavior.
2. Add any behavioral notes not covered by the tool definition to references/behavioral-notes.md.
3. Update the Tool-to-Pipeline-Phase Mapping table in this file.
4. If the tool enables a new workflow pattern, add it to Section 5.
5. Update affected agent definitions if the new tool changes which agents need this skill.