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.

[banderson735]

Hi @FelipeD  - 

Have you had a chance to review this article: Connecting Microsoft 365 Copilot to Qlik's MCP Server with the Agents Toolkit? It walks through creating and deploying declarative agents with the Microsoft 365 Agents Toolkit as one approach for integrating Copilot with Qlik's remote MCP Server. 

[Greg_Taffer]

The OpenAIModelTokenLimit error in Qlik Cloud occurs when data sent to or received from OpenAI via the analytics connector exceeds the maximum context window (token limit) allowed by the specific model, such as GPT-3.5 or GPT-4. This error has been frequently reported in AI-powered tools within the Microsoft/Qlik ecosystem, often indicating that the conversation history or data payload is too large.

[Philipp_Peters]

Hi @nicodataiq,

I'm facing the same issue in Copilot Studio and also in MS Foundry. It's not able to handle measures in the right way. Do you know the reason for that issue? Did you find a fix?

Thanks,
Philipp

[Supersmurf]

Hi all!

I had given up my simple approach with Copilot 365 using a Copilot Studio agent connected to Qlik Cloud MCP. On my first goes I could not ask for anything involving a measure or some kind of calculation. But this week it works with no changes, independent of using GPT or Claude models in the agent.

Next thing was to have Copilot tell a numbers story visually, not only as text. The way we managed this was to instruct the agent to generate visuals as HTML code blocks. Works in Copilot 365, but not as a Teams channel. Makes Copilot a bit better, but Claude is still the rock star! 🤘

Best regards

Jan / Aderian Tallmaker

[Philipp_Peters]

Hi @Supersmurf 

thanks for letting us know, thats great. I'm going to test this in our environment today 🙂 As I wrote, we also tested the MCP with Copilot Studio and Foundry in June and had the same problem - its not possible to work with measures. I created a support ticket for Foundry and they fixed it at the end of june. I was also expecting a solution for Copilot because I thought it was based on the same technology. But no - this was the support's answer: Copilot is a different product and is supported by a separate team. Could you please open a new support ticket with the Microsoft 365 team?

REALLY, MICROSOFT, REALLY??

So I contacted the Copilot support team and informed them about the problem, I also referenced the support ticket for Foundry. I had to explain them everything again, but couldn't finish the problem due to personal absence.

Thats why I really appreciate that it's working now and I dont have to explain the problem again 🤗 Whoever fought the battle - thank you! ⚔️

Best regards,
Philipp

[nicodataiq]

Hi everyone!

As a final contribution to this discussion, I see two viable approaches to building an agent in Copilot Studio that can interact with Qlik's MCP server.

Option 1: Hosting an MCP bridge server

The first approach would be to host our own MCP server that connects to Qlik's MCP server, essentially acting as a proxy or wrapper.

Copilot Studio would send its requests to our MCP bridge, which would then forward them to Qlik's MCP server.

This approach would allow us to properly handle all the required inputs and potentially resolve the current issue with using measures.

Option 2: Connecting Copilot Studio to an external AI agent

The second approach would be to connect Copilot Studio to another AI agent (hosted on AWS, another cloud service, or even a local machine) using asynchronous HTTP requests.

Whenever Copilot needs to interact with Qlik's MCP server, it would delegate the request to this external agent. The agent would handle the MCP communication, execute the necessary operations, and return the results to Copilot Studio.

Both approaches would help overcome the current integration limitations, while keeping Copilot Studio as the main interface for end users.

Hope this helps anyone exploring similar integrations!