Tool Calling sometimes not rendered correctly

[inodb]

What happened?

Occasionally the tool calling will show the raw tool calling codes:

<function_calls>
<invoke name="mcp-clickhouse">
<parameter name="action">query</parameter>
<parameter name="query">

Normally it shows rendered:

Version Information

v0.8.0

Steps to Reproduce

This is a sporadic issue, ask any question and occasionally it'll show the raw sequence rather than the parsed invoke

What browsers are you seeing the problem on?

Chrome

Relevant log output

Added a screenshot below

Screenshots

Broken Rendering

Default Behavior

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

LibreChat does not render tool calling from LLM text, so this is an LLM issue that sometimes happens with Claude.

A common cause for this is when the tool(s) are no longer available, or have been updated, since the conversation first started or since the agent was last configured. It's common for remote MCP servers to update their list of tools, and LC is not yet aware of this change: #7117

It's most common when returning to existing conversation when previous interactions may have the older or no longer available MCP tools.

Claude is also not exempt from hallucinating for a myriad of reasons, and this would qualify as a hallucination.

---

It would help in what context the tool is attempted to be used. Are you using an agent? Try removing the ClickHouse MCP and adding it back to the agent if that's the case.

[inodb]

Thank you! That makes sense. Yes, we are using an agent and our own tweaked version of the clickhouse mcp (https://github.com/cBioPortal/cbioportal-mcp), which mainly just adds some additional system prompt. The above sometimes happens on new conversations without an MCP update. It does call the MCP tool and returns the results properly, but the rendering goes haywire. Sounds like it might be some randomness from the LLM

Not super familiar with how the rendering works. Is it missing some escape sequence? Can try to take a look. Maybe there's a way to hook into the response and handle slight variations from the expected response?