OpenClaw plugin that runs an MCP server, exposing every OpenClaw tool as an individual MCP tool. Instead of proxying entire conversations through a single openclaw_chat tool, RemoteClaw exposes exec, read, write, browser, web_search, memory_search, etc. as separate MCP tools — letting Claude Desktop/Claude.ai call them directly with zero model-inference overhead.

Claude Desktop / Claude.ai
    ↕ MCP Protocol (stdio or SSE)
RemoteClaw MCP Server (OpenClaw plugin)
    ↕ HTTP POST /tools/invoke
OpenClaw Gateway (localhost:18789)
    ↕ tool execution
exec, read, write, browser, web_search, ...

Discovery happens in-process via createOpenClawTools() + createOpenClawCodingTools() — the plugin imports these at runtime from the OpenClaw gateway process to enumerate all available tools with their JSON schemas.

Execution goes through POST /tools/invoke on the gateway — this gets the full policy pipeline, auth, and hooks for free.

src/
├── index.ts              # Plugin entry point (register + service lifecycle)
├── types.ts              # Local type interfaces (AgentTool, McpTool, etc.)
├── tool-discovery.ts     # AgentTool[] → MCP Tool[] conversion
├── tool-invoker.ts       # HTTP client for POST /tools/invoke
├── result-mapper.ts      # OpenClaw result → MCP result mapping
├── mcp-server.ts         # MCP Server factory
└── __tests__/
    ├── result-mapper.test.ts
    ├── tool-invoker.test.ts
    └── mcp-server.integration.test.ts

npm install
npm test          # vitest watch mode
npm run test:run  # single run (31 tests)
npm run typecheck # tsc --noEmit
npm run build     # tsup → dist/index.js

Converts OpenClaw's AgentTool[] (with name, description, parameters) into MCP-compatible tool definitions with inputSchema.

HTTP client that sends POST /tools/invoke requests to the OpenClaw gateway:

{
  "tool": "web_search",
  "args": { "query": "hello" },
  "sessionKey": "main"
}

Supports configurable gateway URL, auth token, session key, and timeout.

Maps between OpenClaw and MCP result formats:

• Text and image content pass through directly
• details field is stripped (MCP doesn't use it)
• {ok: false} responses become {isError: true} MCP results
• Exceptions are caught and mapped to error results

Factory function that wires everything together:

• Registers ListToolsRequestSchema handler for tool enumeration
• Registers CallToolRequestSchema handler for tool execution
• Supports static tool arrays or dynamic functions (re-evaluated on each tools/list)

Registers as an OpenClaw service with start()/stop() lifecycle:

• Dynamically imports tool creators from the OpenClaw gateway process
• Creates ToolInvoker and MCP Server
• Exports all modules for standalone use

Plugin config via openclaw.plugin.json:

Runtime config passed via api.config:

31 tests across 3 suites:

• result-mapper (10 tests) — pure data transformation, zero deps
• tool-invoker (10 tests) — HTTP proxy with mocked fetch
• mcp-server integration (11 tests) — full MCP Client↔Server via InMemoryTransport

Integration tests create a real MCP Client and Server connected in-process, verifying the entire pipeline from client.callTool() through HTTP invocation to result mapping.