MCP Client Skill

Purpose

OpenClaw versions before native MCP client support (e.g. 2026.3.13) do not expose configured MCP servers as agent tools. This skill bridges that gap by spawning MCP server processes directly and speaking JSON-RPC over stdio.

When to use

• User asks to use an MCP server (desktop-commander, playwright, memory, etc.)
• openclaw mcp list returns "unknown command" or config path not found
• Native MCP tool calling is unavailable in the current runtime

When NOT to use

• Native MCP support is already working (openclaw mcp list shows servers)
• The task can be done with existing agent tools (exec, read, edit, browser)

Workflow

1. Discover configured servers

Read ~/.openclaw/openclaw.json (JSON) or ~/.openclaw/config.yaml (YAML) and extract mcp.servers:

{
  "mcp": {
    "servers": {
      "desktop-commander": {
        "command": "node",
        "args": ["/Users/deepak/code/mcps/desktop-commander/dist/index.js"]
      },
      "memory": {
        "command": "npx",
        "args": ["-y", "@modelcontextprotocol/server-memory"],
        "env": { "MEMORY_FILE_PATH": "/Users/deepak/.memory/memory.json" }
      }
    }
  }
}

2. Spawn the server process

Use scripts/mcp-call.js (see below) or inline Node.js to:

1. Spawn the process with command + args
2. Set any env variables
3. Send initialize request (JSON-RPC)
4. Send tools/list to discover available tools
5. Send tools/call to invoke the desired tool
6. Parse the result and return to user
7. Kill the process

3. Handle responses

MCP servers return results as JSON-RPC responses. Tool results are in:

{
  "jsonrpc": "2.0",
  "id": 2,
  "result": {
    "content": [{"type": "text", "text": "..."}],
    "isError": false
  }
}

Extract result.content[0].text for text output, or handle isError: true.

Helper Script

Use scripts/mcp-call.js for reliable MCP communication:

node scripts/mcp-call.js <server-name> <tool-name> '<json-args>'

Examples:

# Read a file via desktop-commander
node scripts/mcp-call.js desktop-commander read_file '{"path":"/Users/deepak/code/mcps/desktop-commander/package.json","length":10}'

# Read memory graph
node scripts/mcp-call.js memory read_graph '{}'

# Search memory nodes
node scripts/mcp-call.js memory search_nodes '{"query":"physics"}'

The script:

• Reads server config from ~/.openclaw/openclaw.json
• Spawns the process, sends initialize + tools/call
• Outputs clean result text (or JSON if --json)
• Exits with code 1 on error

Memory MCP Server Tools

The @modelcontextprotocol/server-memory exposes a knowledge graph:

Security Notes

• Only spawn servers from trusted config (user-owned ~/.openclaw/)
• Respect allowedDirectories in desktop-commander — it will reject paths
• Do not log or echo MCP server env vars (may contain tokens)
• Kill processes after use to avoid zombie MCP servers

Current Limitations

• Stdio transport only (no SSE/streamable-http support in workaround)
• One-shot calls (no persistent MCP connection across agent turns)
• Process spawn overhead per call (~1-2s for feature flag fetch)
• Tool schemas are discovered dynamically, not cached

Future

When OpenClaw natively supports MCP clients (openclaw mcp list works), this skill becomes obsolete. Transition path: remove skill, rely on native bundle-mcp tool profile.