When code interpreter is turned on, the mcp server context is lost

[im360john]

What happened?

latest version, enabled mcp servers to zapier and snowflake with claude 3.7 and code interpreter turned off. works great.

restart service with code interpreter turned on and librechat suddenly loses context of the mcp services.

Version Information

LibreChat v0.7.8

Steps to Reproduce

1. enable zapier and/or snowflake mcpservers with runcode: false, claude 3.7. ask if it is aware of zapier and snowflake -> yes and works.
2. restart service with runcode: true and ask the same mcp awareness question -> no longer aware.

What browsers are you seeing the problem on?

No response

Relevant log output

n/a

Screenshots

No response

Code of Conduct

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

[im360john]

noticed the same thing with artifacts. If I have snowflake mcp context enabled and artifacts enabled, the behavior is to analyze within the mcp first and then any react code output is not displayed within artifacts. if I untoggle the mcp server and tell it to regenerate the visual, artifacts panel kicks in.

[danny-avila]

Hi @im360john thanks for the report.

With the given information, there is not enough to determine a reproducible issue or to help troubleshoot.

Can you depict how and where you have MCP servers enabled? Are you using agents? Can you share the UI of the Agent builder (form) to see what you have configured?

What is your librechat.yaml configued as?

[im360john]

sorry for the confusion, hopefully this update is more clear:

see attached librechat.yaml file. I have turned on artifacts using the GUI settings.

When I have snowflake MCP enabled and ask it to return some results and visualize in a graph, the result is an output of raw code: https://imgur.com/a/yP30RGb

When I have the MCP untoggled and ask to regenerate the visual, the result is in the artifacts window as expected: https://imgur.com/a/80Ki4V6

In native claude interface, it is able to fetch MCP results and also generate artifacts at the same time. So not sure if this is a bug with librechat or an enhancement.

--

Similarly, if i configure mcp with code interpreter (librechat2.yaml) and enable both in the chat option..the llm is unable to realize that mcp is toggled on and fails to make any attempts at using MCP. now the bug seems to be that even if i toggle off code interpreter, the llm still fails to utilize MCP.

librechat2.yaml.txt
librechat.yaml.txt

[danny-avila]

I see, yes. Previously, you would have to go "Settings" and then enable Artifacts. This works with a non-agent endpoint (in other words, when you select Anthropic > Claude 3.7 directly from the top-left dropdown menu).

However, when you select MCP Server(s), this changes the "endpoint" to "Agents" under the hood, which for Artifacts, is only available when you explicitly make an Agent.

I'll address this by also making an "Artifacts" button you can easily toggle on/off, as you won't have to use the Settings to enable it app-wide, only when you want to use it. Something like this: