docs/openclaw.md source ↗

This guide explains the current kmsg + OpenClaw integration model.

Core Model

There are two different interfaces:

MCP, via kmsg mcp-server
real-time streaming, via kmsg watch "<chat>" --json

Today, MCP is still request/response only. It exposes:

kmsg_read
kmsg_send
kmsg_send_image

watch is not an MCP tool. If you want real-time auto-reply, run watch --json as a separate process and feed each event into OpenClaw or a small wrapper around it.

Recommended Architecture

Use two processes:

kmsg mcp-server
kmsg watch "채팅방 이름" --json

The flow is:

watch --json emits a new KakaoTalk event
your supervisor passes that event to OpenClaw
OpenClaw drafts a reply
the reply is sent with kmsg_send through MCP, or kmsg send directly

This separation matters:

MCP handles tool calls cleanly
watch handles low-latency inbound detection
you do not need a streaming MCP extension to operate today

Prerequisites

macOS with KakaoTalk installed
Accessibility permission granted for kmsg
kmsg installed and working

Check first:

kmsg --version
kmsg status

MCP Setup

Run the MCP server:

kmsg mcp-server

The stdio server accepts both MCP Content-Length framing and newline-delimited JSON-RPC input. It replies with the same framing style as the current request, so older OpenClaw-style clients and simpler NDJSON supervisors can use the same binary.

Config example:

{
  "mcpServers": {
    "kmsg": {
      "command": "kmsg",
      "args": ["mcp-server"],
      "env": {
        "KMSG_DEFAULT_DEEP_RECOVERY": "false",
        "KMSG_TRACE_DEFAULT": "false"
      }
    }
  }
}

You can copy:

docs/openclaw.mcp.example.json

Real-Time Watch Setup

Run a dedicated watch process for the chat you want to automate:

kmsg watch "채팅방 이름" --json

watch --json emits one JSON object per detected event on stdout.

Example:

{
  "chat": "홍길동",
  "detected_at": "2026-03-25T10:20:30.123Z",
  "event": "message",
  "message": {
    "author": "홍길동",
    "time_raw": "10:20",
    "body": "새 메시지"
  }
}

Notes:

watch now defaults to 200ms polling
startup uses a short warm-up to avoid backfill
messages earlier than the watch start cutoff are suppressed
--trace-ax logs stay on stderr

Tool Contracts

`kmsg_read`

Input:

{
  "chat": "채팅방 이름",
  "limit": 20,
  "deep_recovery": false,
  "keep_window": false,
  "trace_ax": false
}

Success shape:

{
  "ok": true,
  "chat": "채팅방 이름",
  "fetched_at": "2026-02-26T03:10:10.123Z",
  "count": 20,
  "messages": [
    {
      "author": "홍길동",
      "time_raw": "00:27",
      "body": "밤이 깊었네"
    }
  ],
  "meta": {
    "latency_ms": 1820
  }
}

`kmsg_send`

Input:

{
  "chat": "채팅방 이름",
  "message": "테스트 메시지",
  "confirm": false,
  "deep_recovery": false,
  "keep_window": false,
  "trace_ax": false
}

Notes:

confirm=false or omitted sends immediately
confirm=true does not send and returns CONFIRMATION_REQUIRED

`kmsg_send_image`

Input:

{
  "chat": "채팅방 이름",
  "image_path": "/path/to/image.png",
  "confirm": false,
  "deep_recovery": false,
  "keep_window": false,
  "trace_ax": false
}

Operating Modes

Recommended: Draft Then Approve

watch --json emits a message
OpenClaw drafts a reply
user approves
send with kmsg_send

This is the safest default because KakaoTalk is a personal chat surface and mistakes are expensive.

Minimal send call:

{
  "name": "kmsg_send",
  "arguments": {
    "chat": "홍길동",
    "message": "검토 후 전송합니다.",
    "confirm": false
  }
}

Advanced: Full Auto-Reply

watch --json emits a message
OpenClaw generates a reply automatically
your supervisor immediately calls kmsg_send

Use this only with guardrails. At minimum:

ignore your own messages
restrict to specific chat rooms
add cooldown / loop protection
log every outbound message

Quick Start

Minimal manual setup:

kmsg mcp-server

In another process:

kmsg watch "채팅방 이름" --json

Then wire the watch events into OpenClaw and send replies through MCP kmsg_send.

Troubleshooting

If watch fails:

kmsg watch "채팅방 이름" --json --trace-ax
kmsg inspect --window 0 --depth 20

If MCP fails:

kmsg mcp-server
kmsg status
confirm Accessibility permission and KakaoTalk readiness
if a client sends JSON lines instead of Content-Length frames, keep one JSON-RPC request per line

OpenClaw Integration Guide