Setup Qlik Cloud MCP & LLM Copilot

[FelipeD]

Hi team,

I tried to set up this LLM but currently, to make the test with Swagger on Power Automate to authorize the OAuth session, it indicates that the redirect URI is not registered, and I made this in the before steps.

Some have the same problem or know other steps to work with Qlik Cloud MCP and Copilot.

Thanks

[paulcalvet]

I try it, and Copilot with Qlik doesn't work very well.

I have the error OpenAIModelTokenLimit, I ask to Claude why and this is the the summary :

Is somebody use it or is everybody use Claude ?

---

🔴 The OpenAIModelTokenLimit Error — What's Happening

This error means your conversation has exceeded the maximum token limit allowed by the underlying OpenAI model in your Copilot Studio agent.

The real issue is structural: with an MCP like Qlik, every tool call returns data (schemas, query results, metadata) that accumulates in the conversation context — and that fills up the counter very quickly.

---

📊 Context Window Comparison

  Copilot Studio Claude (direct)

Effective context window	~4k–8k tokens (GPT-3.5/4 depending on config)	200k tokens (Sonnet 4.6)
Underlying model	GPT-3.5 / GPT-4 / GPT-4 Turbo depending on plan	Native Claude Sonnet 4.6
Max MCP payload	Limited — if a real-time connector returns a payload larger than the agent can handle, it throws a blocking error	Up to ~180k tokens of data per exchange
History management	Accumulates → hits limit fast	Natively managed, sliding window
Stateful MCP support	Known issue: when an MCP server holds state, errors appear silently on the Copilot Studio backend — the workaround is to build stateless MCP servers	No restriction on stateful servers
Context control	Very limited for end users	Fully controllable

---

⚠️ Copilot Studio–Specific Limitations with Qlik MCP

1. Token limit too low for Qlik data Qlik MCP can return large JSON objects (datasets, hypercubes, app lists). With GPT-4 at 8,192 tokens, just a few exchanges are enough to saturate the context.

2. Statefulness problem Copilot Studio creates a session ID each time it interacts with a stateful MCP server — this causes silent backend errors. The current recommendation is to build stateless MCP servers for Copilot Studio. A Qlik MCP that maintains an active connection is typically stateful by nature.

3. Connector payload limits If a real-time connector returns a payload larger than the agent can handle, the error is blocking with no graceful fallback.

4. No granular control Unlike Claude directly, you can't easily adjust what enters the context, truncate MCP responses, or manage history compression yourself.

---

✅ Why Claude Is Better Suited for Qlik MCP

• 200k token effective context: you can pass full Qlik hypercubes, app metadata, and multiple exchanges without saturating
• No intermediary layer: Claude calls the MCP directly, without Copilot Studio's constraints (OpenAPI 2.0 requirement, payload limits, etc.)
• Native long-context handling: ideal for multi-step analytical sessions querying Qlik iteratively
• No GPT model dependency: you avoid OpenAIModelTokenLimit errors by design

---

🔧 Workarounds If You Must Keep Copilot Studio

1. Reduce the Qlik payload — configure the MCP to return paginated or truncated responses
2. Reset the conversation regularly (new session) to clear the accumulated context
3. Switch to GPT-4 Turbo in your Copilot Studio settings (128k tokens) if your license allows it
4. Offload heavy lifting to Power Automate — aggregate and summarize Qlik data in a Flow before sending a compact result to the agent

Bottom line: for any serious use of Qlik MCP (iterative queries, large metadata volumes, multi-step analysis), Claude is structurally better suited than Copilot Studio, whose token limits are simply too low for this kind of workload.

[nicodataiq]

Replying to the bug question:

So basically MCP tools expose their arguments so the agent knows what to send. The problem is that the "measures" argument isn't being exposed properly, so Copilot ends up calling the tool with everything except measures and it throws an error. Haven't cracked it yet, if you find something let me know!

[nicodataiq]

Replying to your last post:

I also experienced issues when using GPT models in Copilot, especially with errors that block access to sensitive information and prevent retrieving table data. You can solve this issue by using Claude models.

In my experience, it’s still possible to make an agent work with some effort, particularly for analytics on existing Qlik charts. You can improve results through better prompting and by adding automation flows to the agent. For example, you can retrieve large JSON datasets, process or filter them with a flow, and then return a more compact and usable result back to the agent.