Files
lc-mcp/lc_mcp.py
T

230 lines
10 KiB
Python

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""Limbus Company 语义化操控 MCP server v0.5 (跨平台)。
核心思路: 不让 AI 盲猜坐标。工具层做"文本化 + 语义点击"——
get_state : 截图 + OCR -> 全页文字元素(带坐标)
click_text : 说文字/按钮名, 内部 OCR 定位后点击
wait_for_text / wait_for_text_gone : 等条件出现/消失再回报
平台无关: 截图/输入按 LCP_PLATFORM (win|linux|mac, 默认自动检测) 加载实现。
win : 截图=交互计划任务; 输入=sendinput|interception
linux : 截图=X11 PIL/scrot; 输入=xdotool|uinput(物理)
mac : 截图=screencapture; 输入=cg
OCR 服务: 跑在远端(默认 http://35.212.193.150:8599/), 客户端机器零安装。
协议: MCP stdio (newline-delimited JSON-RPC)。
"""
import os
import sys
import json
import time
import urllib.request
from platform import detect_platform, get_input_backend, get_screenshot_provider
BACKEND_NAME = os.environ.get('INPUT_BACKEND', '').strip().lower()
OCR_URL = os.environ.get('OCR_URL', 'http://35.212.193.150:8599/').strip().rstrip('/') + '/'
MIN_SCORE = float(os.environ.get('OCR_MIN_SCORE', '0.5'))
BASE = os.path.join(os.path.expanduser('~'), 'lc_mcp')
os.makedirs(BASE, exist_ok=True)
def log(msg):
try:
with open(os.path.join(BASE, 'lc_mcp.log'), 'a', encoding='utf-8') as f:
f.write('[%s] %s\n' % (time.strftime('%H:%M:%S'), msg))
except Exception:
pass
def ocr_image(img_path):
"""把本地 png POST 给远端 OCR,返回 items 列表。"""
with open(img_path, 'rb') as f:
body = f.read()
req = urllib.request.Request(OCR_URL, data=body, headers={'Content-Type': 'image/png'})
try:
with urllib.request.urlopen(req, timeout=30) as resp:
data = json.loads(resp.read().decode('utf-8'))
except Exception as e:
return None, 'ocr call fail: %s' % e
if not data.get('ok'):
return None, data.get('error', 'ocr error')
items = [it for it in data.get('items', []) if float(it.get('score', 0)) >= MIN_SCORE]
return items, None
def calc_click(txt, items):
"""在 OCR items 里找包含 txt 的项,返回中心和候选列表。"""
hits = []
for it in items:
if txt.lower() in it['text'].lower():
cx = it['x'] + it['w'] // 2
cy = it['y'] + it['h'] // 2
hits.append({'text': it['text'], 'x': cx, 'y': cy})
return hits
class MCPServer:
def __init__(self, platform, provider, backend):
self.platform = platform
self.provider = provider
self.backend = backend
self.tools = [
{"name": "get_state", "description": "截图并OCR, 返回当前屏幕全部文字元素(含坐标)。AI 决策前先调这个, 不需要自己猜坐标",
"inputSchema": {"type": "object", "properties": {}}},
{"name": "click_text", "description": "语义点击: 给按钮文字(支持子串), 内部OCR定位后点击其中心。例: click_text('开始')",
"inputSchema": {"type": "object", "properties": {
"text": {"type": "string"}, "index": {"type": "number", "default": 0}},
"required": ["text"]}},
{"name": "wait_for_text", "description": "循环截图+OCR, 等到指定文字出现(默认最多30秒)", "inputSchema": {"type": "object", "properties": {
"text": {"type": "string"}, "timeout": {"type": "number", "default": 30}},
"required": ["text"]}},
{"name": "wait_for_text_gone", "description": "循环截图+OCR, 等到指定文字消失(默认最多30秒)", "inputSchema": {"type": "object", "properties": {
"text": {"type": "string"}, "timeout": {"type": "number", "default": 30}},
"required": ["text"]}},
{"name": "screenshot", "description": "截图并返回PNG路径(用于模板/人工查看)", "inputSchema": {"type": "object", "properties": {}}},
{"name": "click", "description": "直接按坐标点击(一般用 click_text 更稳)", "inputSchema": {"type": "object", "properties": {
"x": {"type": "number"}, "y": {"type": "number"},
"button": {"type": "string", "enum": ["left", "right"], "default": "left"}},
"required": ["x", "y"]}},
{"name": "move", "description": "移动鼠标到(x,y)", "inputSchema": {"type": "object", "properties": {
"x": {"type": "number"}, "y": {"type": "number"}}, "required": ["x", "y"]}},
{"name": "keypress", "description": "按一次键(p/enter/esc/space/tab/alt/f4/单字符等)", "inputSchema": {"type": "object", "properties": {
"key": {"type": "string"}}, "required": ["key"]}},
{"name": "type_text", "description": "输入一串文本", "inputSchema": {"type": "object", "properties": {
"text": {"type": "string"}}, "required": ["text"]}},
]
def call(self, name, args):
args = args or {}
if name == 'screenshot':
return self.provider.grab()
if name == 'get_state':
sh = self.provider.grab()
if not sh.get('ok'):
return sh
items, err = ocr_image(sh['path'])
if err:
return {'ok': False, 'error': err}
return {'ok': True, 'screen': sh['path'], 'items': items}
if name == 'click_text':
sh = self.provider.grab()
if not sh.get('ok'):
return sh
items, err = ocr_image(sh['path'])
if err:
return {'ok': False, 'error': err}
hits = calc_click(args.get('text', ''), items or [])
if not hits:
return {'ok': False, 'error': 'text not found: %s' % args.get('text', ''),
'screen': sh['path'], 'items': items}
idx = min(int(args.get('index', 0)), len(hits) - 1)
h = hits[idx]
r = self.backend.click(h['x'], h['y'])
r['on'] = h['text']
return r
if name == 'wait_for_text':
return self._wait(args.get('text', ''), float(args.get('timeout', 30)), True)
if name == 'wait_for_text_gone':
return self._wait(args.get('text', ''), float(args.get('timeout', 30)), False)
if name == 'click':
return self.backend.click(int(args.get('x', 0)), int(args.get('y', 0)), args.get('button', 'left'), 1)
if name == 'move':
return self.backend.move(int(args.get('x', 0)), int(args.get('y', 0)))
if name == 'keypress':
return self.backend.keypress(args.get('key', ''))
if name == 'type_text':
return self.backend.type_text(args.get('text', ''))
return {'ok': False, 'error': 'unknown tool %s' % name}
def _wait(self, txt, timeout, want_present):
deadline = time.time() + timeout
last_items = []
while time.time() < deadline:
sh = self.provider.grab()
if sh.get('ok'):
items, err = ocr_image(sh['path'])
if not err:
last_items = items or []
hits = calc_click(txt, last_items)
if bool(hits) == want_present:
return {'ok': True, 'present': bool(hits),
'hits': hits[:5], 'screen': sh['path']}
time.sleep(1.2)
return {'ok': False, 'error': 'timeout waiting %s (present=%s)' % (txt, want_present),
'last_items': last_items[:10]}
def read_msg():
line = sys.stdin.readline()
if not line:
return None
return json.loads(line.strip())
def send(obj):
sys.stdout.write(json.dumps(obj, ensure_ascii=False) + '\n')
sys.stdout.flush()
def main():
platform = detect_platform()
log('lc_mcp v0.5 platform=%s backend=%s ocr=%s' % (platform, BACKEND_NAME or 'auto', OCR_URL))
backend = None
try:
provider = get_screenshot_provider(platform)
backend = get_input_backend(BACKEND_NAME, platform)
except Exception as e:
log('init fail: %s' % e)
provider = None
backend = None
if backend is None:
log('WARN: input backend unavailable, semantic/click tools disabled')
server = MCPServer(platform, provider, backend)
while True:
try:
msg = read_msg()
if msg is None:
break
except Exception as e:
log('read err %s' % e)
break
method = msg.get('method')
mid = msg.get('id')
if method == 'initialize':
send({"jsonrpc": "2.0", "id": mid, "result": {
"protocolVersion": "2024-11-05",
"capabilities": {"tools": {"listChanged": False}},
"serverInfo": {"name": "lc-mcp", "version": "0.5.0"}}})
elif method == 'notifications/initialized':
continue
elif method == 'ping':
send({"jsonrpc": "2.0", "id": mid, "result": {}})
elif method == 'tools/list':
send({"jsonrpc": "2.0", "id": mid, "result": {"tools": server.tools}})
elif method == 'tools/call':
params = msg.get('params', {})
name = params.get('name')
args = params.get('arguments', {})
log('call %s %s' % (name, args))
if server.backend is None and name not in ('screenshot', 'get_state'):
send({"jsonrpc": "2.0", "id": mid, "result": {"content": [
{"type": "text", "text": json.dumps({'ok': False, 'error': 'input backend unavailable'})}]}})
continue
try:
result = server.call(name, args)
send({"jsonrpc": "2.0", "id": mid, "result": {"content": [
{"type": "text", "text": json.dumps(result, ensure_ascii=False)}]}})
except Exception as e:
log('tool err %s' % e)
send({"jsonrpc": "2.0", "id": mid, "result": {"content": [
{"type": "text", "text": json.dumps({'ok': False, 'error': str(e)})}]}, "isError": True})
else:
if mid is not None:
send({"jsonrpc": "2.0", "id": mid,
"error": {"code": -32601, "message": "method not found"}})
log('lc_mcp exit')
if __name__ == '__main__':
main()