Genie UI MCP Server Connection Issue

[Bihaag_N]

Hi community,

I ran into an issue while connecting a third-party MCP server to Genie's general chat UI and wanted to check if anyone else has seen this.

The connection fails immediately during the initial MCP handshake with:

`TypeError: unhashable type: 'list'`

This happens before any query is even executed.

What's interesting is that the same MCP server works fine in Genie Code. The issue only seems to happen in the general chat UI.

From what I can tell, Genie general chat may be passing the raw MCP response directly into an internal handler without normalising certain fields first. Some third-party MCP servers return list-type values in places like tool manifests or OAuth scopes, and it looks like something downstream expects hashable types instead.

Has anyone else experienced this with external MCP servers, or found a cleaner fix/configuration on the connection side?

[Ashwin_DSA]

Hi @Bihaag_N,

From what I can gather, I couldn't find this exact TypeError: unhashable type: 'list' in the public docs, and I also don’t see any documented connection-side requirement stating that third-party MCP manifests need list fields normalised before Databricks can consume them.

What the public docs do confirm is that Databricks supports external MCP servers via managed proxies, that those servers must use the Streamable HTTP transport, and that they can be tested in Databricks-native surfaces like Genie Code and AI Playground. They also note that external MCP support depends on regions where Model Serving is available. Please take a look at Connect Genie Code to MCP servers and Install an external MCP server.

Given that the same server works in Genie Code but fails immediately during the initial handshake in general chat, this sounds more like a surface-specific product bug than a clean MCP connection misconfiguration. Since the failure happens before any actual tool call, the most useful next step is to capture a HAR from the general chat flow and file it with the initialise / tools-list payloads if you can. As a quick sanity check, I’d also test the same connection in AI Playground, since that’s explicitly documented as a supported way to validate external MCP integrations. Install an external MCP server.

If this answer resolves your question, could you mark it as “Accept as Solution”? That helps other users quickly find the correct fix.