Custom MCP (STDIO transport) extremely slow

[toi500]

Describe the bug

Is this how the Custom MCP (STDIO transport) works now? It is basically spending around 40 seconds to add just 1 action and another 40 seconds to delete it, making it extremely frustrating to use.

chrome_2IDrObzwBd.mp4

To Reproduce

Here is an example:

{ "command": "npx",
"args": [
"@playwright/mcp@latest",
"--headless" ] }

Expected behavior

After the MCP server spins up, add and delete actions without any fuss.

Flowise Version

3.0.3

Additional context

Tested with several MCP servers locally and on production with the same result.

Note: SSE is not affected, working as expected.

[toi500]

@HenryHengZJ I have checked out what is going on and it looks like the root cause of the delay is the re-creation of a child process for the STDIO transport on every single UI refresh.

Custom_MCP.ts

this.inputs = [
    {
        label: 'Available Actions',
        name: 'mcpActions',
        type: 'asyncMultiOptions',
        loadMethod: 'listActions',
        refresh: true              // <-- here
    }
]

I have tested an cache implementation to fix this for good by storing each unique STDIO client in a global map, ensuring the process is created only once per configuration and reused instantly for all subsequent calls.

To make this a complete solution and prevent process leaks over time, we should also implement a destroy() method on the Custom_MCP node. This method would be called automatically when the node is deleted, allowing us to flush the specific client from the cache and terminate its associated child process, avoiding any resource overhead from building up.

Using cache:

[HenryHengZJ]

can you raise a draft PR for the cache implementation you have?

[toi500]

can you raise a draft PR for the cache implementation you have?

Here it is #4809

[toi500]

Fixed here #4897