Experiences with the web integrations in chatbots

[alexanderguentert]

Hey everyone,

I recently discovered your cool MCP and was asked to share my experiences here, so here we go:

I work in the Open Data Team for the City of Zurich, and we were recently asked by a user if we offer an MCP server to search our open government data catalogue using natural language.

We are currently working on a similar topic, namely a chatbot to query our linked data: https://www.melody-bot.ch/

However, it can only query linked data at the moment, so I searched for MCP servers for CKAN (our data catalog). I found two:

• https://github.com/ondics/ckan-mcp-server

• https://github.com/ondata/ckan-mcp-server

Both can be used with different CKAN portals. Since I don't have a local LLM tool installed on my computer, I was happy to find that you offer a hosted trial version: https://ckan-mcp-server.andy-pr.workers.dev/mcp

That's pretty cool, because it allowed me to easily integrate the MCP into the web interfaces of ChatGPT and Le Chat. You even have instructions for ChatGPT, but it was also quite easy with Le Chat (via Intelligence -> Connectors).

Of course, I first experimented with our own data catalog at https://data.stadt-zuerich.ch/. That worked pretty well too. Here are a few examples in English and German.

However, you have to be careful that the tool actually uses the MCP server and does not make other queries. Such as here where it uses the MCP but did not find anything and then used another number from the Internet.

Otherwise, I tried a few of the other CKAN catalogs in Switzerland:

• opendata.swiss also works.
• However, I had problems with https://data.opentransportdata.swiss/. I think their API is not where it usually is with CKAN: https://opentransportdata.swiss/en/cookbook/development-miscellaneous-cookbook/verwendung-des-ckan-api/.
  But I think we could solve that by adding the correct urls here: https://github.com/ondata/ckan-mcp-server/blob/main/src/portals.json

I hope my experiences help you. Otherwise, I wanted to thank you for this great project. It's really cool and I will continue to use it.

Finally, here's a little addition from us: A user has also created an MCP specifically for Open Data Zurich, which also includes a few additional APIs from us: https://github.com/malkreide/zurich-opendata-mcp

[aborruso]

Hi @alexanderguentert, thanks a lot for sharing your experiences!

A quick question about data.opentransportdata.swiss: looking at their CKAN API docs, the API seems to live on a different domain (api.opentransportdata.swiss/ckan-api/) and requires an API key. Since our server only uses public endpoints, this portal would be out of scope.

Do you know if there's a public endpoint that doesn't require registration?

[alexanderguentert]

Hi @aborruso
It is actually true that you need a registration. I did not know that and think it was open a few years ago. I think they've revamped their APIs to bring them all together in one place because they also had some non-open-data APIs that required login credentials.
It would be a bit of a bummer if it only worked with an API key. I'll ask them if there'll be an open version for the non-sensitive data again.

[alexanderguentert]

I asked them about it and they had to put the API behind a login because there were too many requests that were overloading their system. I pointed out that this is not ideal for MCP (and other users). However, they want to focus more on MCP in the future and they know that they will need an open API again for this. Reach out to them at opendata@sbb.ch, if you have questions.

[aborruso]

About the hallucination issue you noticed (when the tool finds nothing, the LLM fills the gap with web data): this is expected LLM behavior, not something we can fully control from the server side. Even if we added a "do not use external sources" note in the tool response itself, the LLM may ignore it — a tool result is just content, not an instruction with the authority of a system prompt.

The most reliable fix is in the system prompt of the app using the MCP server — instructing the model to only use data returned by the tools and not fall back to external sources.

That said, we'll add the note in the tool response anyway — it may help in some cases.

We've added a note about this in the Troubleshooting section of the README.

[alexanderguentert]

Thanks for adding this to the troubleshooting section. I know this behaviour is not unusual (and not in the scope of the MCP server). I just wanted to mention it so that other users know to expect and handle it.

[aborruso]

Thank you. We have added your notes in the 0.4.57 release https://github.com/ondata/ckan-mcp-server/releases/tag/v0.4.57