Load MCP in parallel and cache MCP tools

[mpazik]

The remote MCP servers are painfully slow to load, especially since I am currently in China and coding over VPN.

I noticed that loading sequentially doesn't help. After conducting a few tests, I found that the initial client was the slowest.

Test Results

Config

  "mcp": {
    "context7": {
      "type": "remote",
      "url": "https://mcp.context7.com/sse"
    },
    "coin": {
      "type": "remote",
      "url": "https://mcp.api.coingecko.com/sse"
    },
    "fetch": {
      "type": "remote",
      "url": "https://remote.mcpservers.org/fetch/mcp"
    }
  }

Below are the results from my simple benchmark of three runs:

Original implementation:

• Clients load: 8436ms - 12622ms
• Tools load (per message): 1192ms - 1387ms

Parallel client and tool loading:

• Clients load: 2437ms - 3412ms
• Tools load (per message): 367ms - 569ms

After caching:

• Clients and tools load together
• No additional loading per message
• Drawback: Need to restart opencode on MCP server changes but as it is API I assume it will change very infrequently.