Qlik Sense MCP Server

Model Context Protocol server for Qlik Sense Enterprise. Exposes Qlik's Repository (HTTP) and Engine (WebSocket) APIs as 26 MCP tools so an LLM client can discover apps, inspect data models, build hypercubes, and manage reload tasks through a single uniform interface. In JWT mode the 14 reload-task tools are hidden, since QRS task administration needs certificate auth.

What's in the box

Full list with descriptions: docs/tools.md.

The main analysis call maps onto SQL one-to-one — dimensions is the GROUP BY, measures are the aggregates, sort_by / sort_order are the ORDER BY, limit is the LIMIT:

// engine_create_hypercube — "the 10 clients with the highest GGR"
{
  "app_id": "<app guid>",
  "dimensions": [{"field": "clientid"}],
  "measures":   [{"expression": "Sum(ggr)", "label": "GGR"}],
  "sort_by": "GGR",
  "sort_order": "desc",
  "limit": 10
}

Quick start

uvx qlik-sense-mcp-server

The server starts in Streamable HTTP mode on http://127.0.0.1:8000/mcp. Configure it via environment variables — see docs/configuration.md.

For stdio mode (legacy MCP transport), pass --stdio.

Two authentication modes are supported: client certificate (legacy, full QRS access) and JWT via virtual proxy (per-analyst, no on-disk secrets). See docs/AUTH_JWT.md for the JWT setup.

Documentation

Key facts about the v1.9.0 line

• A wrong query is refused, not answered. Qlik evaluates an unknown field name as an expression worth 0, so a hypercube grouped by a typo came back as a single row holding the grand total — a plausible number with nothing to mark it as wrong. Dimension fields are checked against the data model first (about 10ms), and measures that come back entirely zero, contain SQL, or mention names the model does not have are reported in warnings.

• Objects say which fields they use. get_app_sheet_objects returns fields_used, including fields reached through master measures and the ones inside a filter pane's listboxes — so "what does this sheet work with" is one call.

• Paging is done by Qlik, not after it. App and task listings read /{entity}/table with skip/take and take the total from /{entity}/count, so nothing past the QRS record limit goes missing and total_found is the real total. Field search and field paging likewise happen in Engine — verified on a field with 200,000 distinct values, where the old local scan simply could not see a match.

• A failure is never an empty answer. A QRS 500, a refused connection or an Engine error used to arrive as [], "" or "no schedule", which reads as a tidy, empty Qlik. Every such path now returns an error_category and the original cause.

• Column meanings, not just column names. Fields and tables commented in the load script (COMMENT FIELD / COMMENT TABLE) carry that text into get_app_details as comment, and into get_app_field as field_comment, so the model reads what a column means instead of guessing from its name. Added in 1.7.2.

• Runs on both MCP SDK lines. SDK 2.0 dropped FastMCP; the server now picks MCPServer (2.x) or FastMCP (1.x) at import time, so mcp>=1.1.0,<3.0.0 all work. Both lines are covered by the test suite and were verified end to end against a live Qlik app.

• Ranked queries (top-N) in one call. engine_create_hypercube takes sort_by (a measure label, a measure expression or a dimension field), sort_order (desc / asc) and limit, so "the 10 clients with the highest GGR" is a single request. Before v1.6.0 sorting by a measure silently did nothing — qInterColumnSortOrder was hard-coded to the dimensions, so the server returned the alphabetically first rows instead of the largest ones.

• NULL groups stay out of rankings. Facts with no value for the grouping field collapse into Qlik's "-" row, which often holds a large total and would otherwise take first place in a top-N. It is dropped by default; pass exclude_null_dimensions=false to measure how much data is unattributed.

• Compact, LLM-friendly results. The hypercube response is columns + rows with real numbers, plus grand_total and per-step timings. Pass include_raw_layout=true for the full Qlik layout.

• Failures name the query that failed. Every error reply, timeouts included, echoes tool and request with the exact arguments sent.

• Fewer useless tools in JWT mode. Reload-task administration needs QRS admin rights, so those 14 tools are registered only in certificate mode: 26 tools with a certificate, 12 with a JWT.

• One Qlik session per server. Qlik's per-user limit (5 by default) counts proxy sessions, and in JWT mode one is created by the session bootstrap itself — before any WebSocket. The server therefore bootstraps once and reuses that session for every call; restarting it in a loop is what exhausts the quota, not the number of queries.

• JWT authentication via virtual proxy. Set QLIK_JWT_TOKEN instead of certificate paths and the server will authenticate every Repository and Engine call as the analyst encoded in the token. No certificates or private keys live on the host. The legacy certificate mode is unchanged and still required for full QRS access. Setup guide: docs/AUTH_JWT.md.

• Cached Engine WebSocket connections. Once an app is opened, every subsequent tool call against the same app_id reuses the same WebSocket and the same open document. Switching app_id closes the old document and opens the new one on the same socket. Dropped connections are reopened transparently. Implementation: engine_api.py and docs/architecture.md.

• Streamable HTTP transport by default. The server is a long-lived process; multiple MCP clients can talk to it in parallel. The legacy stdio mode still works behind --stdio.

• tool_call_seconds is injected as the first key of every tool response — wall-clock time of the call in milliseconds. Use it to spot slow tools.

• Hard hypercube limits. engine_create_hypercube rejects requests with max_rows > 5000 or columns * max_rows > 9900 immediately, with a structured error and a hint pointing at set-analysis or top-N patterns. Qlik Engine itself returns error 7009 calc-pages-too-large for any single page over 10000 cells.

• Single timeout knob. QLIK_WS_TIMEOUT (default 180.0 seconds) controls both the WebSocket handshake and every Engine API call.

Requirements

• Python 3.12 (the package is built and tested against this version; see pyproject.toml)
• Qlik Sense Enterprise (Repository on port 4242, Engine on port 4747 — the standard ports)
• Client certificate, private key and root CA from the Qlik Sense node
• Network access from the host running this server to Qlik

Disclaimer

This project is an independent, community-built integration. It is NOT affiliated with, endorsed by, sponsored by, or supported by Qlik Technologies Inc., QlikTech International AB, or any other Qlik entity. "Qlik", "Qlik Sense", "QlikView" and all related product names are trademarks of their respective owners.

All information about Qlik Sense APIs, port allocations, error codes, protocol behavior and usage patterns used in this project was obtained exclusively from publicly available sources — the Qlik Developer Portal (help.qlik.com, qlik.dev), the Qlik Community forums, and other public documentation. No proprietary, confidential or reverse-engineered material is used.

License

MIT © 2025-2026 Stanislav Chernov