openclaw - 💡(How to fix) Fix Feature request: per-agent MCP server scoping [1 comments, 2 participants]

Problem

When multiple agents are configured on a single gateway, every MCP server defined in mcp.servers is spawned once per agent — regardless of whether that agent needs those tools.

For example, with 12 agents and 10 MCP servers, the gateway spawns 120 processes even though only 4-5 agents actually use those MCP tools. The remaining agents (e.g. simple chat/social agents) get processes they'll never call.

On a 32GB instance with 10 server types and 12 agents, the unnecessary MCP processes consume ~3GB of RAM.

Proposed solution

Add an optional agents field to mcp.servers.* config that accepts an array of agent IDs (or glob patterns):

{
  "mcp": {
    "servers": {
      "my-erp": {
        "command": "python3",
        "args": ["/path/to/server.py"],
        "agents": ["admin-agent", "work-group-*"]
      },
      "lightweight-tool": {
        "command": "node",
        "args": ["server.js"]
        // no agents field = available to all (current default behaviour)
      }
    }
  }
}

When agents is set, the gateway would only spawn that MCP server for matching agents. When omitted, current behaviour is preserved (spawn for all).

Alternatives considered

• tools.deny on agents: hides MCP tools from the prompt but processes still spawn and consume RAM
• Separate gateway instances: works but adds operational complexity for what should be a config-level concern

Impact

This would significantly reduce memory usage for multi-agent deployments where not all agents need the same MCP tools.