diff --git a/.gitignore b/.gitignore index a122b6c..ef27ec3 100644 --- a/.gitignore +++ b/.gitignore @@ -27,6 +27,9 @@ ssh public/local-ai.json .env +# MCP 服务生成产物(工程 ZIP / 配音,运行时数据,勿提交) +mcp-out/ + # 数据质量检查的生成报告(每次运行都会变,不入库) data-quality/reports/ diff --git a/CHANGELOG.md b/CHANGELOG.md index c0ae99b..872fed4 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -6,6 +6,28 @@ ## 2026-11-11 +### ✨ 素材面板菜单 + +**素材面板新增「下载更多素材」直达入口** + +- 上传素材右侧的下拉菜单(在线录音 / 文字转语音)新增第三项「下载更多素材」,点击在新窗口打开素材下载文档站 `https://doc.88531.cn/docs/mrfx/mrfx-1g99qum57n5qk` +- 文案入语言表:`zhCN.my_download_more` / `enUS.my_download_more`,图标用 `Download`(@icon-park/react),菜单项渲染与既有项同构 + +### 🤖 MCP 接入 + +**新增 `unicut-mcp` 服务:AI 客户端直接调用米剪能力(Model Context Protocol)** + +- 新增 `mcp/` 服务(Python FastMCP,Streamable HTTP):抓取文章 `grab_url`(与前端 `extractFromHtml` 同规则)、口播文案 `generate_script`(OpenAI 兼容网关)、配音 `tts` / `list_voices`(unicut-tts)、工程生成 `create_video_project`、文件夹模式工程读写 `list/read/write_project`、自检 `service_status` +- `create_video_project` 产出与米剪导出同构的工程 ZIP(`project.json + materials/{type}/`,过 `checkNdata` 九键门禁):图片校验对齐前端(≥160×120 / ≤10MB)、有配音自动铺满时长、文案按字数加权切字幕,元素结构照抄官方 demo 工程真实字段 +- nginx 同源反代 `https://<站点>/mcp`(SSE 关缓冲长连接)与 `https://<站点>/mcp-out/*`(工程 ZIP 下载);compose 新增 `unicut-mcp` 服务,密钥走项目 `.env`(`MCP_LLM_KEY` 等,不入库) +- 文档:`mcp/README.md`(工具明细 + Claude Desktop / Claude Code / Cursor / Hermes 接入配置 + docker run / compose 两种部署),主 README「AI」章新增 MCP 小节与更新表行 +- 版本号 `1.0.0 → 1.0.1`(`package.json` / README 发布 tag 同步;授权中心基线仍为 1.0.0,客户端上报 1.0.1 判定「已是最新」) + +### 🐛 修复(/grab/ 抓取) + +- 公众号图片防盗链:`/grab/` 向上游清空 `Referer`(带异源 Referer 时 `mmbiz.qpic.cn` 只给 140×140 降级缩略图,无 Referer 才给原图,实测 2KB vs 47KB) +- 正文根节点优先 `#js_content`:微信文章页没有 `article`/`main`,旧选择器会退化到整页 `body` 混进账号栏/留言噪音 + ### ✨ 新功能 **AI 一键出片 · 素材三选一(图片 / Word·PDF 文档 / 网址)** diff --git a/README.md b/README.md index de53ace..e8e2f0d 100644 --- a/README.md +++ b/README.md @@ -53,9 +53,11 @@ | 模板导入修复 | 修复"导入第一个模板后,再导入会把浏览器拖崩"的问题,支持连续换模板 | | 品牌与首页 | 产品名改为米剪,首页新增导航站 / 开源地址入口,移除收费栏目与账号类外链 | | 接入授权中心 | 版本号、公告、检查更新(授权中心不可达时静默降级,绝不拦功能) | +| MCP 接入 | 新增 `unicut-mcp` 服务:Claude / Cursor / Hermes 等 AI 客户端经 `https://<站点>/mcp` 调用抓取文章、生成口播、edge-tts 配音、生成米剪工程 ZIP | ### 最近更新(2026-11-11) +- **素材面板「下载更多素材」入口**:上传素材旁的下拉菜单(在线录音 / 文字转语音)新增第三项,点击新窗口直达素材下载文档站 https://doc.88531.cn/docs/mrfx/mrfx-1g99qum57n5qk; - **AI 一键出片素材三选一**:图片 / Word·PDF 文档(自动提取正文与配图,扫描件有提示)/ 网址自动抓取正文和配图(同源 `/grab/` 反代,拒绝内网回环目标);纯文字素材自动补标题封面; - **纯文本模型也能出片**:出片卡片新增提示「只需纯文本模型即可使用;模型支持多模态(视觉)就能自动生成出片」;未配视觉模型时剥掉图片块、按图片编号循环分配画面,纯文本主模型不再被顶回错误; - **断网 lint 门禁修复**:`external_refs` 20 → 0(PDF.js 命名空间入册、示例域名清除、官方自有域名注明理由入白名单),构建门禁恢复全绿; @@ -188,6 +190,20 @@ UNICUT 是一款面向创作者、开发者与企业团队的在线视频编辑 > [!IMPORTANT] > AI 对话、AI 一键出片、AI 生图 / 生视频默认接入官方 AI 网关(`ai.88531.cn`),在编辑器右上角「配置模型」粘贴一个 API Key 即可解锁,对话 `mimo-v2.6-flash`、视觉 `deepseek-v4-flash-vision-exp`、生图 `flux-schnell`、生视频 `inkling` 均已预填。配音(本机 edge-tts)、抠图、字幕、导出等能力不消耗任何外部额度。 +### MCP:让 AI 客户端直接调用米剪 + +米剪内置 [MCP](https://modelcontextprotocol.io)(Model Context Protocol)服务 `unicut-mcp`,站点同源端点 `https://<站点>/mcp`。接入后,Claude Desktop、Claude Code、Cursor、Hermes 等 AI 客户端可以把米剪当工具用: + +| 工具 | 能力 | +| --- | --- | +| `grab_url` | 抓网页正文与配图(公众号文章可用,规则与编辑器内抓取一致) | +| `generate_script` | 大模型生成短视频口播文案 | +| `tts` / `list_voices` | edge-tts 多音色配音(本机免费) | +| `create_video_project` | 图片 + 文案 + 配音 → 生成米剪工程 ZIP,米剪「工程文件」一键导入即可出片 | +| `list_projects` / `read_project` / `write_project` | 工作文件夹模式下的工程读写 | + +典型工作流:*AI 抓文章 → 写口播 → 配音 → 生成工程 ZIP → 你在米剪导入*。部署、客户端接入配置与工具明细见 [mcp/README.md](./mcp/README.md)。 + ## 🧰 技术实现 米剪采用 React + TypeScript 构建编辑器界面与业务层,通过 MobX 管理复杂编辑状态;画布由 PixiJS / WebGL 实时渲染,音视频解码、处理和导出由 Web Worker、FFmpeg WebAssembly、MediaBunny 等能力协同完成。素材文件可写入用户授权的本地目录,工程元数据和文件句柄通过 Dexie / IndexedDB 管理。 @@ -214,7 +230,7 @@ docker run -d --name unicut -p 3006:80 -p 3005:443 \ gitea.88531.cn/jinghui/unicut:latest ``` -打开 `http://服务器IP:3006` 即用。Registry 地址 `gitea.88531.cn` 免登录可拉;镜像与源码构建产物完全一致,每次发版同步更新 `latest` 与 `v1.0.0` 两个 tag。 +打开 `http://服务器IP:3006` 即用。Registry 地址 `gitea.88531.cn` 免登录可拉;镜像与源码构建产物完全一致,每次发版同步更新 `latest` 与 `v1.0.1` 两个 tag。 ### 方式二:Docker Compose部署(源码 compose 构建) @@ -246,6 +262,23 @@ services: volumes: - ../unicut-tts/audio:/data/audio + # MCP 服务(Model Context Protocol):AI 客户端经 https://<站点>/mcp 调用米剪能力 + # (抓取正文配图 / 大模型口播文案 / edge-tts 配音 / 生成米剪工程 ZIP)。 + # 密钥与站点地址从项目 .env 注入(MCP_LLM_KEY 等,勿提交仓库)。 + unicut-mcp: + build: ./mcp + container_name: unicut-mcp + restart: always + environment: + - MCP_LLM_BASE=${MCP_LLM_BASE:-http://192.168.68.19:3099/v1} + - MCP_LLM_KEY=${MCP_LLM_KEY:-} + - MCP_LLM_MODEL=${MCP_LLM_MODEL:-mimo-v2.6-flash} + - MCP_PUBLIC_URL=${MCP_PUBLIC_URL:-https://127.0.0.1:3005} + - MCP_PROJECTS_DIR=${MCP_PROJECTS_DIR:-} + volumes: + # 生成的工程 ZIP / 配音 mp3 落盘(nginx /mcp-out/ 提供下载) + - ./mcp-out:/data/out + volumes: certs: ``` @@ -254,7 +287,7 @@ volumes: git clone https://gitee.com/jinghui111/unicut.git cd unicut docker compose up -d unicut # 只启动编辑器 -docker compose up -d # 全量(需 ../unicut-tts 源码) +docker compose up -d # 全量(含配音 unicut-tts 与 MCP unicut-mcp) ``` 首次构建包含前端打包,约需几分钟。完成后浏览器访问: @@ -266,6 +299,7 @@ http://服务器IP:3006 - 端口映射:`3006 → 80`(HTTP)、`3005 → 443`(HTTPS,自签证书,首次访问需接受一次证书警告); - 首次进入按提示选择一个本地文件夹并授权,工程与素材即保存在该文件夹; - 可选:在 `.env` 里配置 `HUB_KEY` / `HUB_PROJECT_ID` 接入授权中心(公告 / 版本检查),不配置不影响任何功能; +- 可选:在 `.env` 里配置 `MCP_LLM_KEY`(OpenAI 兼容网关 Key)启用 MCP 的文案生成,AI 客户端连接 `https://<站点>/mcp`,工具明细见 [mcp/README.md](./mcp/README.md);不配置则除大模型类工具外均可使用; - 完整版另含配音服务 `unicut-tts`(本机 edge-tts),需单独部署该服务后用 `docker compose up -d` 全量启动;只用编辑器按上面的 `up -d unicut` 即可。 ### 方式三:本地开发 diff --git a/docker-compose.yml b/docker-compose.yml index 0ee8d05..ed4ac1d 100644 --- a/docker-compose.yml +++ b/docker-compose.yml @@ -23,5 +23,22 @@ services: volumes: - ../unicut-tts/audio:/data/audio + # MCP 服务(Model Context Protocol):AI 客户端经 https://<站点>/mcp 调用米剪能力 + # (抓取正文配图 / 大模型口播文案 / edge-tts 配音 / 生成米剪工程 ZIP)。 + # 密钥与站点地址从项目 .env 注入(MCP_LLM_KEY 等,勿提交仓库)。 + unicut-mcp: + build: ./mcp + container_name: unicut-mcp + restart: always + environment: + - MCP_LLM_BASE=${MCP_LLM_BASE:-http://192.168.68.19:3099/v1} + - MCP_LLM_KEY=${MCP_LLM_KEY:-} + - MCP_LLM_MODEL=${MCP_LLM_MODEL:-mimo-v2.6-flash} + - MCP_PUBLIC_URL=${MCP_PUBLIC_URL:-https://127.0.0.1:3005} + - MCP_PROJECTS_DIR=${MCP_PROJECTS_DIR:-} + volumes: + # 生成的工程 ZIP / 配音 mp3 落盘(nginx /mcp-out/ 提供下载) + - ./mcp-out:/data/out + volumes: certs: diff --git a/mcp/Dockerfile b/mcp/Dockerfile new file mode 100644 index 0000000..0209a75 --- /dev/null +++ b/mcp/Dockerfile @@ -0,0 +1,17 @@ +# 米剪 MCP 服务:Streamable HTTP 端点 /mcp(8765),由 unicut nginx 同源反代 +FROM docker.1ms.run/library/python:3.11-slim + +ENV PYTHONUNBUFFERED=1 \ + PIP_INDEX_URL=https://pypi.tuna.tsinghua.edu.cn/simple \ + PIP_DISABLE_PIP_VERSION_CHECK=1 + +WORKDIR /app + +COPY requirements.txt . +RUN pip install --no-cache-dir -r requirements.txt + +COPY server.py . + +EXPOSE 8765 + +CMD ["python", "server.py"] diff --git a/mcp/README.md b/mcp/README.md new file mode 100644 index 0000000..c99e487 --- /dev/null +++ b/mcp/README.md @@ -0,0 +1,107 @@ +# 米剪 MCP 服务(unicut-mcp) + +让 AI 客户端以 [Model Context Protocol](https://modelcontextprotocol.io) 工具的形式调用米剪的创作能力:抓文章、写口播、配语音、生成米剪工程,人在米剪里打开就能剪。 + +``` +AI(Claude / Cursor / Hermes …) + │ MCP(Streamable HTTP) + ▼ +https://<米剪站点>/mcp ──► unicut-mcp 容器(本目录) + ├─ grab_url 抓网页正文与配图(公众号文章可用) + ├─ generate_script 大模型生成口播文案(OpenAI 兼容网关) + ├─ tts / list_voices edge-tts 配音(unicut-tts 容器) + ├─ create_video_project 生成米剪工程 ZIP → 米剪「工程文件」导入 + └─ list/read/write_project 工作文件夹(文件夹模式)工程读写 +``` + +## 工具一览 + +| 工具 | 作用 | 典型返回 | +| --- | --- | --- | +| `service_status` | 自检抓取/配音/大模型/工程目录可用性 | 各依赖状态 | +| `grab_url(url)` | 抓取网页正文与配图(规则与前端 `extractFromHtml` 一致:`#js_content` 优先、懒加载属性、12000 字截断) | `{title, text, images[]}` | +| `generate_script(source, requirement?, style?, target_seconds?, paragraph_count?)` | 大模型写口播文案,严格 JSON 输出 | `{title, paragraphs[], text}` | +| `list_voices()` | edge-tts 音色列表 | `{count, voices[]}` | +| `tts(text, voice?, rate?, pitch?, volume?)` | 合成配音,音频落盘并给出可播放 URL | `{url, path, duration}` | +| `create_video_project(title, images, script?, audio?, width?, height?, …)` | 下载素材 → 组装 MovieData → 打包工程 ZIP(`project.json + materials/`,与米剪导出格式同构,过 `checkNdata` 九键门禁) | `{zip_path, zip_url, duration, captions, elements}` | +| `list_projects()` / `read_project(id)` / `write_project(project)` | 工作文件夹模式下直接读写 `projects/*.json`(需配置 `MCP_PROJECTS_DIR`) | 工程 JSON | + +`create_video_project` 细节: + +- 图片校验与前端一致:≥160×120、≤10MB、识别失败自动跳过并在 `warnings` 里说明; +- 有配音时图片自动拉伸平铺铺满配音时长,无配音时每图 `image_duration`(默认 3 秒); +- `script` 按 `\n` 分段成字幕(单段超 24 字自动按句拆),按字数加权分配时间; +- 字幕/图片/音频元素结构照抄官方 demo 工程(`public/assets/demo/test-app.json`)真实字段。 + +## 连接方式 + +**端点**:`https://<米剪站点>/mcp`(Streamable HTTP;本机默认 `https://127.0.0.1:3005/mcp`,自签证书需客户端允许 insecure)。局域网内也可直连容器端口 `http://<主机>:8765/mcp`(默认只在 compose 网络内,未映射宿主端口)。 + +**Claude Desktop**(`claude_desktop_config.json`): + +```json +{ + "mcpServers": { + "unicut": { + "command": "npx", + "args": ["mcp-remote", "https://127.0.0.1:3005/mcp", "--allow-self-signed"] + } + } +} +``` + +**Claude Code / Cursor**: + +```bash +claude mcp add unicut --transport http https://127.0.0.1:3005/mcp +# Cursor:~/.cursor/mcp.json → {"mcpServers":{"unicut":{"url":"https://127.0.0.1:3005/mcp"}}} +``` + +**Hermes**(`~/.hermes` 或项目 `config.yaml`): + +```yaml +mcp: + servers: + unicut: + transport: http + url: https://127.0.0.1:3005/mcp +``` + +## 部署 + +compose 已内置 `unicut-mcp` 服务(本仓库 `docker-compose.yml`),密钥走项目 `.env`: + +```bash +# .env(勿提交) +MCP_LLM_BASE=http://192.168.68.19:3099/v1 # OpenAI 兼容 /v1 网关 +MCP_LLM_KEY=sk-xxxx # 网关 Key +MCP_LLM_MODEL=mimo-v2.6-flash +MCP_PUBLIC_URL=https://127.0.0.1:3005 # 对外站点地址(生成 zip_url / 音频 URL 用) +MCP_PROJECTS_DIR= # 可选:米剪「选择文件夹」的同一目录(文件夹模式) +``` + +```bash +docker compose up -d --build unicut-mcp +``` + +单独运行(不走 nginx): + +```bash +docker build -t unicut-mcp ./mcp +docker run -d --name unicut-mcp --restart always \ + -e MCP_LLM_KEY=sk-xxxx -e MCP_LLM_BASE=http://192.168.68.19:3099/v1 \ + -p 8765:8765 -v "$PWD/mcp-out:/data/out" unicut-mcp +``` + +> 只有 `service_status` / `grab_url` / `generate_script` 需要网络依赖;`tts` 依赖同 compose 的 `unicut-tts`;不配置 `.env` 时除大模型类工具外均可使用。 + +## 验证 + +```bash +# 协议握手 +curl -sk https://127.0.0.1:3005/mcp -H 'Content-Type: application/json' \ + -H 'Accept: application/json, text/event-stream' \ + -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"curl","version":"0"}}}' +``` + +产物下载:`https://<站点>/mcp-out/<工程名>.zip`;健康检查:经 nginx 用 `https://<站点>/mcp-health`,直连容器用 `http://<容器>:8765/health`(`/mcp/health` 不是端点,会被 SPA 回退成首页 HTML)。 diff --git a/mcp/requirements.txt b/mcp/requirements.txt new file mode 100644 index 0000000..2484a14 --- /dev/null +++ b/mcp/requirements.txt @@ -0,0 +1,5 @@ +mcp>=1.9.0,<2 +httpx>=0.27.0 +beautifulsoup4>=4.12.0 +pillow>=10.0.0 +mutagen>=1.47.0 diff --git a/mcp/server.py b/mcp/server.py new file mode 100644 index 0000000..e5d67f7 --- /dev/null +++ b/mcp/server.py @@ -0,0 +1,838 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- +""" +米剪(Unicut)MCP 服务 +====================== + +让 AI 客户端(Claude Desktop / Claude Code / Cursor / Hermes 等)以 MCP 工具的形式 +调用米剪的创作能力: + + - grab_url 抓取网页正文与配图(公众号文章等,复用前端 extractFromHtml 规则) + - generate_script 大模型生成短视频口播文案 + - tts / list_voices edge-tts 配音合成(unicut-tts) + - create_video_project 组装米剪工程 ZIP(project.json + materials/),在米剪「工程文件」一键导入 + - list_projects / read_project 工作文件夹(文件夹模式)下的工程读写 + - service_status 各依赖服务可用性自检 + +传输:Streamable HTTP,端点 /mcp(容器 8765,nginx 同源反代 https://<站点>/mcp)。 +工程格式:与米剪 projectExport.ts 导出的 ZIP 完全同构—— + project.json MovieData(title/poster/width/height/background/transitions/ + captions/elements/resouces 九键齐全,过 checkNdata 门禁) + materials/{type}/{文件名} 素材相对路径 +""" + +from __future__ import annotations + +import io +import ipaddress +import json +import os +import re +import socket +import time +import unicodedata +import uuid +import zipfile +from pathlib import Path +from urllib.parse import urljoin, urlparse + +import httpx +from bs4 import BeautifulSoup +from mutagen.mp3 import MP3 +from PIL import Image + +from mcp.server.fastmcp import FastMCP + +# ---------------------------------------------------------------- 环境配置 + +UNICUT_URL = os.getenv("UNICUT_URL", "http://unicut") # nginx 80(回环到 443) +TTS_URL = os.getenv("TTS_URL", "http://unicut-tts:8000") +LLM_BASE = os.getenv("MCP_LLM_BASE", "http://192.168.68.19:3099/v1").rstrip("/") +LLM_KEY = os.getenv("MCP_LLM_KEY", "") +LLM_MODEL = os.getenv("MCP_LLM_MODEL", "mimo-v2.6-flash") +PUBLIC_URL = os.getenv("MCP_PUBLIC_URL", "https://127.0.0.1:3005").rstrip("/") +PROJECTS_DIR = os.getenv("MCP_PROJECTS_DIR", "") # 文件夹模式工作目录(可空) +OUT_DIR = Path(os.getenv("MCP_OUT_DIR", "/data/out")) + +IMG_MIN_W, IMG_MIN_H = 160, 120 # 与前端 sourceImport.ts 常量一致 +IMG_MAX_BYTES = 10 * 1024 * 1024 +HTTP_HEADERS = { + # 无 Referer:公众号图床 mmbiz.qpic.cn 带异源 Referer 时只给降级缩略图(实测) + "User-Agent": ( + "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 " + "(KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36" + ), + "Accept": "*/*", + "Accept-Language": "zh-CN,zh;q=0.9", +} + +mcp = FastMCP( + "unicut", + instructions=( + "米剪(Unicut)在线视频剪辑器的 MCP 服务。" + "典型工作流:grab_url 抓文章 → generate_script 写口播 → tts 配音 → " + "create_video_project 生成米剪工程 ZIP → 用户在米剪「工程文件」导入即可出片。" + ), + host="0.0.0.0", + port=8765, +) + + +def _client(timeout: float = 30.0) -> httpx.AsyncClient: + # trust_env=False:不读宿主代理变量,容器网络行为可预期 + return httpx.AsyncClient(timeout=timeout, follow_redirects=True, trust_env=False) + + +# ---------------------------------------------------------------- 安全与小工具 + +def assert_public_url(url: str) -> str: + """拒绝非 http(s) 与内网/回环目标(与 nginx /grab/ 的断言同规则,防 SSRF 打内网)""" + p = urlparse(url) + if p.scheme not in ("http", "https"): + raise ValueError(f"只支持 http/https URL:{url}") + host = p.hostname + if not host: + raise ValueError(f"URL 缺少主机名:{url}") + if host == "localhost" or host.endswith(".local"): + raise ValueError(f"拒绝内网/回环目标:{host}") + try: + infos = socket.getaddrinfo(host, None, proto=socket.IPPROTO_TCP) + except socket.gaierror as e: + raise ValueError(f"域名解析失败:{host}({e})") + for info in infos: + ip = ipaddress.ip_address(info[4][0]) + if ip.is_private or ip.is_loopback or ip.is_link_local or ip.is_reserved or ip.is_multicast: + raise ValueError(f"拒绝内网地址目标:{host} → {ip}") + return str(p) + + +def _uid() -> str: + return uuid.uuid4().hex[:10] + + +def _slug(name: str) -> str: + """文件名安全化:保留中英文数字,替换掉路径/控制字符""" + s = unicodedata.normalize("NFKC", name or "project") + s = re.sub(r'[\\/:*?"<>|\r\n\t]', "-", s).strip(" .") + return (s or "project")[:60] + + +def _ext_for(content_type: str, url: str, default: str = "bin") -> str: + ct = (content_type or "").split(";")[0].strip().lower() + table = { + "image/jpeg": "jpg", "image/jpg": "jpg", "image/png": "png", + "image/webp": "webp", "image/gif": "gif", "audio/mpeg": "mp3", + "audio/mp3": "mp3", "audio/wav": "wav", "audio/x-wav": "wav", + "audio/mp4": "m4a", "audio/aac": "aac", + } + if ct in table: + return table[ct] + suffix = Path(urlparse(url).path).suffix.lstrip(".").lower() + return suffix or default + + +def _is_local_path(s: str) -> bool: + return s.startswith("/") and not s.startswith("//") + + +# ---------------------------------------------------------------- 抓取(对齐前端 extractFromHtml) + +_BLOCK_TAGS = [ + "p", "h1", "h2", "h3", "h4", "h5", "h6", "li", "blockquote", "pre", + "td", "th", "figcaption", "tr", "div", "br", +] +_NOISE_TAGS = [ + "script", "style", "noscript", "svg", "iframe", "canvas", "form", + "nav", "footer", "header", "aside", "button", +] + + +def extract_from_html(html: str, base_url: str) -> dict: + """标题 / 正文 / 配图(相对转绝对、懒加载属性、去重)——与前端规则一致""" + soup = BeautifulSoup(html, "html.parser") + + og = soup.find("meta", attrs={"property": "og:title"}) + title = ((og.get("content") if og else None) or soup.title.get_text() if soup.title else "").strip() + + for el in soup.find_all(_NOISE_TAGS): + el.decompose() + + root = ( + soup.find(id="js_content") # 公众号正文(微信页无 article/main) + or soup.find("article") + or soup.find("main") + or soup.find(attrs={"role": "main"}) + or soup.body + or soup + ) + + # 图片:og:image 优先 + img 常见懒加载属性 + seen: set[str] = set() + images: list[str] = [] + + def push(v: str | None) -> None: + if not v: + return + s = v.strip() + if not s or s.startswith("data:") or s.startswith("blob:"): + return + if re.search(r"\.svg(\?|#|$)", s, re.I): + return + try: + absu = urljoin(base_url, s) + except Exception: + return + if not re.match(r"^https?:", absu, re.I) or absu in seen: + return + seen.add(absu) + images.append(absu) + + og_img = soup.find("meta", attrs={"property": "og:image"}) + push(og_img.get("content") if og_img else None) + for img in root.find_all("img"): + for attr in ("src", "data-src", "data-original", "data-lazy-src", "data-actualsrc"): + push(img.get(attr)) + + # 正文:块级边界补换行 → textContent → 压空白(照抄前端清洗链) + for el in root.find_all(_BLOCK_TAGS): + el.insert(0, "\n") + el.append("\n") + text = ( + root.get_text() + .replace("\r", "") + .replace("\t", " ") + .replace(" ", " ") + ) + lines = [seg.strip() for seg in text.split("\n")] + kept: list[str] = [] + for i, seg in enumerate(lines): + if seg or (i > 0 and lines[i - 1]): + kept.append(seg) + text = re.sub(r"\n{3,}", "\n\n", "\n".join(kept)).strip()[:12000] + + return {"title": title, "text": text, "images": images[:20]} + + +async def fetch_bytes(url: str, *, allow_local_file: bool = True, timeout: float = 25.0) -> tuple[bytes, str]: + """下载字节。url 可为公网 URL,或(默认允许)服务器本地文件路径。""" + if _is_local_path(url): + if not allow_local_file: + raise ValueError("此上下文不允许本地路径") + p = Path(url) + if not p.is_file(): + raise ValueError(f"本地文件不存在:{url}") + return p.read_bytes(), "" + assert_public_url(url) + async with _client(timeout) as c: + r = await c.get(url, headers=HTTP_HEADERS) + r.raise_for_status() + return r.content, r.headers.get("content-type", "") + + +# ---------------------------------------------------------------- 大模型 + +async def _llm_chat(messages: list[dict], *, temperature: float = 0.7, max_tokens: int = 8000) -> str: + if not LLM_KEY: + raise ValueError("未配置 MCP_LLM_KEY(compose .env),无法调用大模型") + async with _client(timeout=180.0) as c: + r = await c.post( + f"{LLM_BASE}/chat/completions", + headers={"Authorization": f"Bearer {LLM_KEY}"}, + json={ + "model": LLM_MODEL, + "messages": messages, + "temperature": temperature, + "max_tokens": max_tokens, + }, + ) + if r.status_code >= 400: + raise ValueError(f"大模型网关 {r.status_code}: {r.text[:300]}") + data = r.json() + return data["choices"][0]["message"]["content"] + + +def _parse_json_block(s: str) -> dict | None: + s = s.strip() + m = re.search(r"\{.*\}", s, re.S) + if not m: + return None + try: + return json.loads(m.group(0)) + except json.JSONDecodeError: + # 容忍 ```json 围栏 / 尾逗号 + cleaned = re.sub(r"^```(?:json)?|```$", "", m.group(0), flags=re.M).strip() + cleaned = re.sub(r",\s*([}\]])", r"\1", cleaned) + try: + return json.loads(cleaned) + except json.JSONDecodeError: + return None + + +# ---------------------------------------------------------------- 工程组装(MovieData) + +def _split_captions(script: str, total: float, audio_dur: float | None) -> list[dict]: + """口播文案 → 字幕段(按字符数加权切分时间轴;有配音时铺满配音时长)""" + raw = [ln.strip() for ln in re.split(r"\n+", script) if ln.strip()] + lines: list[str] = [] + for ln in raw: + if len(ln) <= 24: + lines.append(ln) + else: # 长段按句切(保留标点) + parts = [s for s in re.split(r"(?<=[。!?!?;;])", ln) if s.strip()] + lines.extend([p.strip() for p in parts] if parts else [ln]) + if not lines: + return [] + + timeline = audio_dur if audio_dur else total + if timeline <= 0: + return [] + weights = [max(len(x), 1) for x in lines] + s = sum(weights) + caps: list[dict] = [] + t = 0.0 + for i, (line, w) in enumerate(zip(lines, weights)): + dur = timeline * w / s + if i == len(lines) - 1: + dur = max(timeline - t, 0.3) # 末段补到结尾,避免浮点欠尾 + caps.append(_caption_element(line, t, max(dur, 0.3))) + t += dur + return caps + + +def _caption_element(text: str, start: float, dur: float, height: int = 1920, fill: str = "rgba(255, 255, 255, 1)") -> dict: + """字幕元素:结构照抄 demo/test-app.json 的真实字幕(字号缩放随画布高度)""" + scale = round(height / 1920, 4) or 1 + return { + "id": _uid(), + "name": text, + "text": text, + "textStyle": { + "radius": [0, 0, 0, 0], + "type": "newTextStyle", + "direction": "horizontal", + "fillType": "normal", + "fill": fill, + "padding": [5, 5, 5, 5], + "fontSize": 200, + "letterSpacing": 0, + "lineHeight": 200, + "align": "center", + "fontFamily": "抖音美好体", + "stroke": {"color": "rgba(51, 51, 51, 1)", "lineJoin": "round", "lineWidth": 18}, + "fontFamilyURL": "/assets/fonts/抖音美好体.otf", + }, + "style": { + "x": 540, + "y": round(height * 0.77, 2), + "alpha": 1, + "width": 392.96, + "height": 67.2, + }, + "type": "caption", + "trackIndex": 0.5, + "startTime": round(start, 3), + "duration": round(dur, 3), + "textScale": round(0.32 * scale, 4), + "scale": 1, + "templateEnable": True, + "sceneIndex": 1, + } + + +def _image_element(resource_id: str, name: str, start: float, dur: float, w: int, h: int, idx: int) -> dict: + """图片元素:字段集对齐 demo 视频元素(同一 BaseElement 血统)+ ImageElement 特有 matting""" + return { + "id": _uid(), + "name": name, + "type": "image", + "trackIndex": 0, + "resourceId": resource_id, + "startTime": round(start, 3), + "duration": round(dur, 3), + "style": {"width": w, "height": h, "x": w / 2, "y": h / 2}, # x/y 是中心点(demo 实证) + "blendMode": 0, + "flipx": False, + "scale": 1, + "filters": [], + "matting": { + "enabled": False, + "color": "#00FF00", + "similarityThreshold": 0.5, + "emergence": 0.5, + }, + "sceneIndex": idx + 1, + "templateEnable": False, + } + + +def _audio_element(resource_id: str, dur: float, volume: float, name: str = "配音.mp3") -> dict: + return { + "id": _uid(), + "name": name, + "type": "audio", + "trackIndex": 1, + "resourceId": resource_id, + "startTime": 0, + "duration": round(dur, 3), + "muted": False, + "volume": volume, + "speed": 1, + "curveSpeed": False, + "clipTime": 0, + "fadeInTime": 0, + "fadeOutTime": 0, + } + + +def build_movie_data( + title: str, + w: int, + h: int, + images: list[dict], # [{res_id, zip_rel, name, iw, ih, mime, size}] + audio: dict | None, # {res_id, zip_rel, name, dur, mime, size} + caption_lines: str, + with_caption: bool, + caption_fill: str, + voice_volume: float, + image_duration: float, +) -> dict: + now = int(time.time() * 1000) + resouces: list[dict] = [] + elements: list[dict] = [] + + n = len(images) + if audio: + total = float(audio["dur"]) + span = total / n if n else 0.0 + else: + span = float(image_duration) + total = n * span + + for i, im in enumerate(images): + resouces.append({ + "id": im["res_id"], + "name": im["name"], + "type": "image", + "url": im["zip_rel"], + "originId": im["res_id"], + "thumb": "", + "styleSize": {"width": im["iw"], "height": im["ih"]}, + "fileType": im["mime"], + "from": "user", + "mustFetch": False, + "attrs": {"width": im["iw"], "height": im["ih"], "size": im["size"], "fileType": im["mime"]}, + }) + elements.append(_image_element(im["res_id"], im["name"], i * span, span, w, h, i)) + + if audio: + resouces.append({ + "id": audio["res_id"], + "name": audio["name"], + "type": "audio", + "url": audio["zip_rel"], + "originId": audio["res_id"], + "thumb": "", + "styleSize": {"width": 0, "height": 0}, + "fileType": audio["mime"], + "from": "user", + "mustFetch": False, + "duration": audio["dur"], + "attrs": {"size": audio["size"], "fileType": audio["mime"], "duration": audio["dur"]}, + }) + elements.append(_audio_element(audio["res_id"], audio["dur"], voice_volume, audio["name"])) + + captions = _split_captions(caption_lines, total, audio["dur"] if audio else None) if with_caption else [] + + # checkNdata 门禁九键 + demo 同款元数据 + return { + "title": title, + "createTime": now, + "updateTime": now, + "poster": "", + "version": "1.1.0", + "width": w, + "height": h, + "duration": round(total, 3), + "background": {}, + "_hideLock": {}, + "transitions": [], + "cameras": [], + "captions": captions, + "elements": elements, + "resouces": resouces, + } + + +def write_project_zip(movie: dict, files: dict[str, bytes], zip_path: Path) -> None: + """project.json + materials/{type}/{file} —— 与 projectExport.ts 导出同构""" + zip_path.parent.mkdir(parents=True, exist_ok=True) + with zipfile.ZipFile(zip_path, "w", zipfile.ZIP_DEFLATED) as z: + z.writestr("project.json", json.dumps(movie, ensure_ascii=False)) + for rel, blob in files.items(): + z.writestr(rel, blob) + + +# ---------------------------------------------------------------- MCP 工具 + +@mcp.tool() +async def service_status() -> dict: + """自检:抓取反代 / 配音服务 / 大模型网关 / 工程目录 的可用性""" + out: dict[str, object] = {"time": time.strftime("%Y-%m-%d %H:%M:%S")} + + try: + async with _client(8) as c: + r = await c.get(f"{UNICUT_URL}/") + out["unicut_nginx"] = f"OK({r.status_code})" + except Exception as e: + out["unicut_nginx"] = f"FAIL: {e}" + + try: + async with _client(8) as c: + r = await c.get(f"{TTS_URL}/voices") + v = r.json() + n = len(v) if isinstance(v, list) else len(v.get("voices", [])) if isinstance(v, dict) else 0 + out["tts"] = f"OK, {n} 个音色" + except Exception as e: + out["tts"] = f"FAIL: {e}" + + out["llm"] = { + "base": LLM_BASE, + "model": LLM_MODEL, + "key": f"已配置({len(LLM_KEY)}位)" if LLM_KEY else "未配置 MCP_LLM_KEY", + } + out["projects_dir"] = PROJECTS_DIR or "未配置(文件夹模式工具不可用)" + out["out_dir"] = str(OUT_DIR) + out["public_url"] = PUBLIC_URL + return out + + +@mcp.tool() +async def grab_url(url: str, max_images: int = 20) -> dict: + """抓取网页正文与配图(支持公众号 mp.weixin.qq.com 文章)。 + + 返回 {ok, url, title, text(≤12000字), images[](绝对地址, 最多 max_images 张)}。 + 图片只是候选地址;create_video_project 下载时会再校验尺寸(≥160x120)与大小(≤10MB)。 + """ + assert_public_url(url) + async with _client(30) as c: + r = await c.get(url, headers=HTTP_HEADERS) + r.raise_for_status() + final_url = str(r.url) + ctype = r.headers.get("content-type", "") + if "html" not in ctype and "xml" not in ctype: + # 非 HTML(如直接是文本/JSON):原样返回文本 + return {"ok": True, "url": final_url, "title": final_url, "text": r.text[:12000], "images": []} + ext = extract_from_html(r.text, final_url) + return { + "ok": True, + "url": final_url, + "title": ext["title"], + "text": ext["text"], + "images": ext["images"][: max(0, min(int(max_images), 20))], + } + + +@mcp.tool() +async def generate_script( + source: str, + requirement: str = "", + style: str = "口播解说", + target_seconds: int = 60, + paragraph_count: int = 6, +) -> dict: + """大模型生成短视频口播文案。 + + source:素材文本(抓取的文章正文)或主题描述;requirement:额外要求(受众/卖点/关键词)。 + 返回 {ok, title, paragraphs[], text};paragraphs 每段一句,直接对应字幕与时间轴分段。 + """ + sys_prompt = ( + "你是短视频口播文案编剧。只输出一个 JSON 对象,不要任何解释、不要 markdown 围栏:\n" + '{"title": "视频标题", "paragraphs": ["第一段口播", "第二段口播", ...]}\n' + "要求:口语化、有钩子、每段 1~2 句话适合做字幕;总字数 ≈ 目标秒数 × 4.5 字。" + ) + user_prompt = ( + f"风格:{style}\n目标时长:约 {target_seconds} 秒\n段落数:约 {paragraph_count} 段\n" + ) + if requirement: + user_prompt += f"补充要求:{requirement}\n" + user_prompt += f"\n素材:\n{source[:8000]}" + + content = await _llm_chat([ + {"role": "system", "content": sys_prompt}, + {"role": "user", "content": user_prompt}, + ]) + data = _parse_json_block(content) + if data and isinstance(data.get("paragraphs"), list): + paras = [str(p).strip() for p in data["paragraphs"] if str(p).strip()] + title = str(data.get("title") or "").strip() + else: # 兜底:按行切 + paras = [ln.strip() for ln in content.splitlines() if ln.strip()] + title = "" + if not paras: + raise ValueError(f"模型未返回可用文案:{content[:300]}") + return {"ok": True, "title": title, "paragraphs": paras, "text": "\n".join(paras)} + + +@mcp.tool() +async def list_voices() -> dict: + """列出配音(edge-tts)可用音色 → {ok, count, voices[]}""" + async with _client(15) as c: + r = await c.get(f"{TTS_URL}/voices") + r.raise_for_status() + v = r.json() + voices = v if isinstance(v, list) else v.get("voices", []) + return {"ok": True, "count": len(voices), "voices": voices} + + +@mcp.tool() +async def tts( + text: str, + voice: str = "zh-CN-XiaoxiaoNeural", + rate: str = "+0%", + pitch: str = "+0Hz", + volume: str = "+0%", +) -> dict: + """edge-tts 配音合成。返回 {ok, url(站点可直接播放), path(服务器文件), duration(秒), voice}。 + text 建议 ≤2000 字;传给 create_video_project 的 audio 用 path。 + """ + if not text.strip(): + raise ValueError("text 不能为空") + async with _client(180) as c: + r = await c.post( + f"{TTS_URL}/tts", + json={"text": text, "voice": voice, "rate": rate, "pitch": pitch, "volume": volume}, + ) + if r.status_code >= 400: + raise ValueError(f"TTS {r.status_code}: {r.text[:300]}") + j = r.json() + if not isinstance(j, dict) or not j.get("url"): + raise ValueError(f"TTS 返回异常:{str(j)[:300]}") + u = str(j["url"]) + internal = TTS_URL + (u[len("/localtts"):] if u.startswith("/localtts") else u) + ar = await c.get(internal) + ar.raise_for_status() + blob = ar.content + if not blob: + raise ValueError("TTS 音频为空") + + audio_dir = OUT_DIR / "audio" + audio_dir.mkdir(parents=True, exist_ok=True) + name = Path(urlparse(u).path).name or f"{_uid()}.mp3" + dest = audio_dir / name + dest.write_bytes(blob) + try: + dur = float(MP3(io.BytesIO(blob)).info.length) + except Exception: + dur = 0.0 + public = PUBLIC_URL + (u if u.startswith("/") else "/" + u) + return {"ok": True, "url": public, "path": str(dest), "duration": round(dur, 3), "voice": voice} + + +@mcp.tool() +async def create_video_project( + title: str, + images: list[str] | None = None, + script: str = "", + audio: str = "", + width: int = 1080, + height: int = 1920, + image_duration: float = 3.0, + caption: bool = True, + caption_fill: str = "rgba(255, 255, 255, 1)", + voice_volume: float = 1.0, +) -> dict: + """生成米剪工程 ZIP(project.json + materials/),在米剪顶部「工程文件」导入即可出片。 + + images:图片 URL 或服务器本地路径(tts 工具产出的 path 也行);script:口播文案→自动做字幕 + (按\\n分段,单段过长自动按句拆);audio:配音 URL/path(通常来自 tts 工具)。 + 有配音时图片自动拉伸铺满配音时长,否则每图 image_duration 秒。 + 返回 {ok, zip_path, zip_url(可直接下载), id, duration, images_used, captions, elements}。 + """ + images = images or [] + warnings: list[str] = [] + + # ---- 图片收集与校验 + im_items: list[dict] = [] + files: dict[str, bytes] = {} + for raw in images: + if len(im_items) >= 30: + warnings.append("图片超过 30 张,多余已丢弃") + break + try: + blob, ctype = await fetch_bytes(raw) + except Exception as e: + warnings.append(f"图片下载失败(跳过){raw[:80]}: {e}") + continue + if len(blob) > IMG_MAX_BYTES: + warnings.append(f"图片超 10MB(跳过){raw[:80]}") + continue + try: + im = Image.open(io.BytesIO(blob)) + iw, ih = im.size + except Exception as e: + warnings.append(f"图片无法识别(跳过){raw[:80]}: {e}") + continue + if iw < IMG_MIN_W or ih < IMG_MIN_H: + warnings.append(f"图片太小 {iw}x{ih}(跳过){raw[:80]}") + continue + ext = _ext_for(ctype, raw, "jpg") + if ext not in ("jpg", "png", "webp", "gif"): + ext = "jpg" if im.format in ("JPEG", "PNG", "WEBP", "GIF") else "jpg" + fname = f"{uuid.uuid4().hex}.{ext}" + rel = f"materials/image/{fname}" + files[rel] = blob + im_items.append({ + "res_id": _uid(), + "name": Path(urlparse(raw).path).name or fname, + "zip_rel": rel, + "iw": iw, "ih": ih, + "mime": f"image/{'jpeg' if ext == 'jpg' else ext}", + "size": len(blob), + }) + + # ---- 配音收集 + aud: dict | None = None + if audio: + try: + blob, ctype = await fetch_bytes(audio) + ext = _ext_for(ctype, audio, "mp3") + fname = f"voice_{uuid.uuid4().hex[:8]}.{ext}" + rel = f"materials/audio/{fname}" + files[rel] = blob + try: + dur = float(MP3(io.BytesIO(blob)).info.length) + except Exception: + dur = 0.0 + if dur <= 0: + warnings.append("配音时长无法识别,时间轴将按图片时长排布") + aud = None + else: + aud = { + "res_id": _uid(), + "name": Path(urlparse(audio).path).name or fname, + "zip_rel": rel, + "dur": dur, + "mime": f"audio/{'mpeg' if ext == 'mp3' else ext}", + "size": len(blob), + } + except Exception as e: + warnings.append(f"配音获取失败(忽略): {e}") + + if not im_items and not aud: + raise ValueError("至少需要 1 张可用图片或 1 段配音") + + # ---- 组装 MovieData + movie = build_movie_data( + title=title, w=int(width), h=int(height), + images=im_items, audio=aud, + caption_lines=script, with_caption=bool(caption and script.strip()), + caption_fill=caption_fill, voice_volume=float(voice_volume), + image_duration=float(image_duration), + ) + + # ---- 落盘 ZIP + ts = time.strftime("%Y%m%d-%H%M%S") + zip_name = f"{_slug(title)}-{ts}.zip" + zip_path = OUT_DIR / zip_name + write_project_zip(movie, files, zip_path) + + # ---- 门禁自检:checkNdata 九键 + required = ["title", "poster", "width", "height", "background", "transitions", "captions", "elements", "resouces"] + missing = [k for k in required if k not in movie] + if missing: + raise ValueError(f"工程缺少必需键:{missing}") + + return { + "ok": True, + "id": f"mcp-{ts}-{_uid()[:6]}", + "zip_path": str(zip_path), + "zip_url": f"{PUBLIC_URL}/mcp-out/{zip_name}", + "duration": movie["duration"], + "size": zip_path.stat().st_size, + "images_used": len(im_items), + "has_audio": bool(aud), + "captions": len(movie["captions"]), + "elements": len(movie["elements"]), + "warnings": warnings, + "import_hint": "下载该 ZIP → 米剪顶部「工程文件」→ 导入(需先选择过工作文件夹)", + } + + +@mcp.tool() +async def list_projects() -> dict: + """列出工作文件夹(文件夹模式 PROJECTS_DIR)里的米剪工程 → {ok, dir, projects[]}""" + if not PROJECTS_DIR: + return {"ok": False, "error": "未配置 MCP_PROJECTS_DIR(compose .env)。文件夹模式需把它指到米剪里选择的同一目录"} + pdir = Path(PROJECTS_DIR) / "projects" + if not pdir.is_dir(): + return {"ok": True, "dir": str(pdir), "projects": []} + items = [] + for f in sorted(pdir.glob("*.json")): + try: + d = json.loads(f.read_text(encoding="utf-8")) + items.append({ + "file": f.name, + "id": d.get("id") or f.stem, + "title": d.get("title") or f.stem, + "width": d.get("width"), + "height": d.get("height"), + "updatedAt": d.get("updatedAt"), + }) + except Exception as e: + items.append({"file": f.name, "error": str(e)}) + return {"ok": True, "dir": str(pdir), "projects": items} + + +@mcp.tool() +async def read_project(project_id: str) -> dict: + """读取工作文件夹里的工程 JSON(folder 形态 {id,title,width,height,data})""" + if not PROJECTS_DIR: + return {"ok": False, "error": "未配置 MCP_PROJECTS_DIR"} + f = Path(PROJECTS_DIR) / "projects" / f"{_slug(project_id)}.json" + if not f.is_file(): + return {"ok": False, "error": f"工程不存在:{f}"} + data = json.loads(f.read_text(encoding="utf-8")) + return {"ok": True, "path": str(f), "project": data} + + +@mcp.tool() +async def write_project(project: dict) -> dict: + """写入/更新工作文件夹工程(folder 形态:{id,title,width,height,data})。谨慎调用:会覆盖同 id 文件""" + if not PROJECTS_DIR: + return {"ok": False, "error": "未配置 MCP_PROJECTS_DIR"} + if not isinstance(project, dict) or not project.get("data"): + return {"ok": False, "error": "需要 {id,title,width,height,data(MovieData)} 结构"} + pid = _slug(str(project.get("id") or _uid())) + project["id"] = pid + project["updatedAt"] = int(time.time() * 1000) + pdir = Path(PROJECTS_DIR) / "projects" + pdir.mkdir(parents=True, exist_ok=True) + f = pdir / f"{pid}.json" + f.write_text(json.dumps(project, ensure_ascii=False), encoding="utf-8") + return {"ok": True, "path": str(f), "id": pid} + + +# ---------------------------------------------------------------- 静态产物与健康检查路由 + +try: + from starlette.responses import FileResponse, JSONResponse + + @mcp.custom_route("/health", methods=["GET"]) + async def _health(_req): + return JSONResponse({"ok": True, "service": "unicut-mcp"}) + + @mcp.custom_route("/out/{name}", methods=["GET"]) + async def _out(req): + name = Path(req.path_params["name"]).name # 防目录穿越 + f = OUT_DIR / name + if not f.is_file(): + return JSONResponse({"ok": False, "error": "not found"}, status_code=404) + return FileResponse(f, filename=name) + +except AttributeError: # 老版本 SDK 无 custom_route,静默降级 + pass + + +if __name__ == "__main__": + OUT_DIR.mkdir(parents=True, exist_ok=True) + mcp.run(transport="streamable-http") diff --git a/nginx.conf b/nginx.conf index feabf5d..cd884b3 100644 --- a/nginx.conf +++ b/nginx.conf @@ -116,6 +116,9 @@ server { proxy_set_header Host $proxy_host; proxy_set_header Connection ""; proxy_set_header X-Grab-Url ""; + # 公众号图片防盗链:带异源 Referer 时 mmbiz.qpic.cn 只给降级缩略图 + # (实测 2KB/140x140,无 Referer 则 47KB/原尺寸),向上游清空该头 + proxy_set_header Referer ""; proxy_http_version 1.1; proxy_read_timeout 20s; proxy_send_timeout 20s; @@ -143,6 +146,46 @@ server { client_max_body_size 8m; } + # MCP 服务(Model Context Protocol,unicut-mcp 容器)同源反代: + # AI 客户端(Claude Desktop / Claude Code / Cursor / Hermes 等)用 + # https://<站点>/mcp + # 连接即可把米剪当工具用(抓取文章 / 生成口播文案 / edge-tts 配音 / 生成工程 ZIP)。 + # Streamable HTTP 走 SSE 流:必须关缓冲、保持 HTTP/1.1 长连接、拉长读超时。 + # 注:/mcp-out 前缀比 /mcp 更长,nginx 最长前缀匹配规则会优先后者,不冲突。 + location /mcp { + proxy_pass http://unicut-mcp:8765; + proxy_http_version 1.1; + proxy_set_header Host $host; + proxy_set_header X-Real-IP $remote_addr; + proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for; + proxy_set_header X-Forwarded-Proto $scheme; + proxy_set_header Connection ""; + proxy_buffering off; + proxy_request_buffering off; + proxy_cache off; + proxy_read_timeout 600s; + proxy_send_timeout 600s; + client_max_body_size 20m; + } + + # MCP 生成产物下载(工程 ZIP、配音 mp3),配合米剪「工程文件」导入 + location /mcp-out/ { + proxy_pass http://unicut-mcp:8765/out/; + proxy_http_version 1.1; + proxy_set_header Host $host; + proxy_set_header X-Real-IP $remote_addr; + proxy_read_timeout 300s; + client_max_body_size 100m; + } + + # MCP 健康检查(容器 /health;注意 /mcp/health 不是端点,SPA 会回退成首页 HTML) + location = /mcp-health { + proxy_pass http://unicut-mcp:8765/health; + proxy_http_version 1.1; + proxy_set_header Host $host; + proxy_read_timeout 10s; + } + # UNICUT 在线服务反代(账号/素材/AI 等在线能力), 对应官方 dev server 的代理配置 location /api/ { proxy_pass https://video.h5ds.com; diff --git a/package.json b/package.json index 87ac5b3..899ac5c 100644 --- a/package.json +++ b/package.json @@ -1,7 +1,7 @@ { "name": "aacut", "private": true, - "version": "1.0.0", + "version": "1.0.1", "scripts": { "dev": "vite", "build": "yarn clean && node scripts/patch-sdk-offline.mjs && yarn build:worker && vite build", diff --git a/src/language/enUS.ts b/src/language/enUS.ts index c08f783..2465610 100644 --- a/src/language/enUS.ts +++ b/src/language/enUS.ts @@ -80,6 +80,7 @@ export const enUS = { my_upload: 'Upload', my_upload_mobile_link: 'Upload Mobile', my_upload_refresh: 'Refresh List', + my_download_more: 'Download More Assets', my_tab_material: 'Current Materials', my_tab_all_material: 'All Materials', my_tab_no_material: 'No materials in this project. Please upload first.', diff --git a/src/language/zhCN.ts b/src/language/zhCN.ts index 180bef2..2f8ce7a 100644 --- a/src/language/zhCN.ts +++ b/src/language/zhCN.ts @@ -91,6 +91,7 @@ export const zhCN = { my_upload: '上传素材', my_upload_mobile_link: '手机上传', my_upload_refresh: '刷新列表', + my_download_more: '下载更多素材', my_tab_material: '当前素材', my_tab_all_material: '全部素材', my_tab_no_material: '该项目暂无素材,请先上传', diff --git a/src/pages/editor/components/sources/my/List.tsx b/src/pages/editor/components/sources/my/List.tsx index 6fe15ca..e803e58 100644 --- a/src/pages/editor/components/sources/my/List.tsx +++ b/src/pages/editor/components/sources/my/List.tsx @@ -12,7 +12,7 @@ import { Pagination, Empty, } from '@douyinfe/semi-ui'; -import { Voice, Left, VoiceInput, Upload as UploadIcon, DeleteFive, Close, Refresh } from '@icon-park/react'; +import { Voice, Left, VoiceInput, Upload as UploadIcon, DeleteFive, Close, Refresh, Download } from '@icon-park/react'; import { useCallback, useEffect, useReducer, useRef, useState } from 'react'; import { getUploadBeforeData, isAPNG } from '@pages/editor/tools/uploadBeforeData'; import { pubsub, util } from '@utils/index'; @@ -497,6 +497,14 @@ export default function List(props: IProps) { setShowType('tts'); }, }, + { + node: 'item', + icon: , + name: language.val('my_download_more'), + onClick: () => { + window.open('https://doc.88531.cn/docs/mrfx/mrfx-1g99qum57n5qk', '_blank'); + }, + }, ]} trigger="click" position="bottomRight" diff --git a/src/utils/sourceImport.ts b/src/utils/sourceImport.ts index b9ddd96..a6144b1 100644 --- a/src/utils/sourceImport.ts +++ b/src/utils/sourceImport.ts @@ -313,6 +313,9 @@ export function extractFromHtml( .forEach(el => el.remove()); const root = + // 公众号文章正文在 #js_content(页面其余部分是账号栏/留言/分享等噪音), + // 必须优先于 article/main 选择器——微信页两者都没有,否则会退化到整页 body + dom.querySelector('#js_content') || dom.querySelector('article') || dom.querySelector('main') || dom.querySelector('[role="main"]') || diff --git a/stats.html b/stats.html index 777dba9..3632f11 100644 --- a/stats.html +++ b/stats.html @@ -4929,7 +4929,7 @@ var drawChart = (function (exports) {