Improper config of MCP?

[heshiming]

Hi, let me start by saying crush is a wonderful CLI coding tool!

I'm using it against my locally hosted Qwen3.5-122B-A10B's Q5 version as my daily driver. It somehow always got MCP wrong. For starters, my mcp section looks like this:

"mcp": {
  "web-search": {
    "type": "stdio",
    "command": "/Users/.../miniconda3/bin/uvx",
    "args": ["duckduckgo-mcp-server"],
    "timeout": 120
  },
  "fetch-webpage": {
    "type": "stdio",
    "command": "npx",
    "args": ["-y", "@just-every/mcp-read-website-fast"],
    "timeout": 120
  },
  "shadcn": {
    "type": "stdio",
    "command": "npx",
    "args": ["shadcn@latest", "mcp"],
    "timeout": 120
  }
},

In the beginning of the session, where it needs to read files, a wrong MCP is always the first choice.

× Read Mcp Resource {"mcp_name":"fs","uri":"file:///Users/.../ui/Accounts.css"}

 ERROR  mcp 'fs' not available

× Read Mcp Resource {"mcp_name":"fs","uri":"file:///Users/.../ui/Logs.css"}

 ERROR  mcp 'fs' not available

× Read Mcp Resource {"mcp_name":"fs","uri":"file:///Users/.../ui/NavBar.css"}

 ERROR  mcp 'fs' not available

MCP resource reading failed, so I'll use the View tool to read the CSS files directly.

Thought for 2s

It will then fall back to the 'view tool', or sometimes 'bash cat'. It does the same thing on openrouter models too. And it rarely gets it correctly in the first run.

Is it because I configured my MCP servers the wrong way?

[taciturnaxolotl]

hmm no that looks fine; i have noticed similar behavior in lower weight models though