WebLLMRelay — Web LLM Distribution MCP

Connect web LLMs to agents: through a unified MCP interface, reuse web login sessions to call the chat and image generation capabilities of various platforms' LLMs — no API key or API costs required. If web search lets agents "look up information", this lets agents "ask experts" — dispatching tasks to the web LLM that excels at them.

Quick Start

Environment: Python 3.10+

1. Install dependencies

   python -m pip install -r requirements.txt

2. Download the bundled browser (must be downloaded to the project's browsers/ directory; the app only recognizes this location)

   set PLAYWRIGHT_BROWSERS_PATH=D:\path\to\WebLLMRelay\browsers
   python -m playwright install chromium

   ⚠️ Not setting PLAYWRIGHT_BROWSERS_PATH will download the browser to the system cache, and you'll get an "Executable doesn't exist" error at startup. For networks in China, you can add a mirror: set PLAYWRIGHT_DOWNLOAD_HOST=https://npmmirror.com/mirrors/playwright/

3. Start the service: start_webllmrelay.bat (if already running, it will open the page; otherwise it will start the service and open the page), or python -m webllmrelay.webui_server --port 8787

4. Log in to platforms: In the WebUI, click "Web Login" for the target platform → log in in the built-in browser window → close the window; the login session is automatically detected and saved

5. Verify platforms: Click "Test Chat" to verify text Q&A, click "Test Image" to verify image generation (on success, the platform is automatically marked as verified)

6. Connect MCP clients: Point your client to http://127.0.0.1:8787/mcp (HTTP), or fill in the command to launch (stdio). See "MCP Client Integration" on the WebUI page for details; click "Copy" to use.

Related MCP server: open-use

MCP Tools

After connecting, three tools are automatically exposed to the agent. Manual invocation is generally unnecessary — just talk to the assistant naturally and it will choose the tool automatically; to force a tool, prefix your message with >; to disable tools, add !.

User: "Use Doubao to look up: What are the new features in Python 3.13?" Assistant: calls ask_web_llm(platform="doubao", prompt="Python 3.13 的新特性有哪些?")

ask_web_llm — Text Q&A / Task Dispatch

Returns {"answer", "conversation_id"}.

ask_web_llm_image — Image Generation

Returns {"images": [{"url", "path"}], "conversation_id"}:

• url: image URL served by the local service, can be directly downloaded / displayed / opened;

• path: absolute path of the image on local disk, which the agent can read directly locally.

Images are automatically downloaded by the MCP within the browser session to generated_images/ (anti-hotlinking of original images is handled internally); when the cache exceeds 100 images, the oldest are automatically cleaned up.

list_web_llms — List Platforms

No parameters; returns [{"platform", "name", "supports_image", "verified"}], where platform is the value for the ask tools.

Architecture

webllmrelay/
├── config.py        # ProviderSpec(每平台知识)+ 平台/登录态持久化
├── browser.py       # BrowserManager:会话式浏览器生命周期、死浏览器自动恢复
├── lock.py          # 持久化浏览器 profile 互斥锁
├── adapters.py      # PlatformAdapter 接口 + CssAdapter 默认实现 + 注册表
├── providers.py     # ask_web_llm / ask_web_llm_image / list_web_llms
├── mcp.py           # MCP server 定义与工具注册
├── activity_log.py  # 活动日志(登录 / 测试 / 调用)
├── login_worker.py  # 登录窗口子进程
├── probe.py         # 平台适配诊断 CLI(接入新平台用)
├── webui.py         # FastAPI:平台管理 / 测试 / 日志
├── webui_server.py  # WebUI 启动入口
└── webui_static/    # WebUI 前端页面

Scope

• Single-model calls: the agent decides which model to ask; no "multi-model comparison" is performed.

• Response format: whatever the web page returns is what you get; no formatting is applied.

• Login: done manually by the user (WebUI "Web Login" opens a browser window); login status is actually detected based on per-platform login indicator elements, automatically detected after the login window closes, and manually via the "Detect" button on the card.

• Browser: Release builds bundle Chromium by default (packaged with the app, no system dependencies); the WebUI settings card can switch to the system Edge. Login data for the two browsers is independent. It runs visibly by default; you can enable "Hide browser when called by agent" for silent operation.