Replies: 1 comment
|
这个方向很扎实——"探索 → 验证 → 保存 → 复用"的学习回路在当前 Harness 生态里确实是空白:Harness 自带的 web 能力是 几个从 Harness 插件接缝角度的观察,供你取舍: 1. WebMCP 工具该挂在哪一层 2. 与 skill 系统的分工 3. 验证证据要可重放 4. 上下文成本 5. 插件形态 想看到你把验证回路再往前推一步——尤其是"网站改版后自动发现工具失效并触发重新验证"的检测,那会是 WebMCP 生命周期里最难的环节,也是最有复用价值的部分。 |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
中文 · English
中文
我在 DeepDeck(基于 DeepSeek Harness 的社区非官方桌面客户端) 中实现了一个支持 WebMCP 的浏览器。它可以调用网站已有的工具,也可以让 Agent 为没有 WebMCP 的网站构建工具。
用户描述任务目标,Agent 自己尝试网站操作、检查结果,并把验证过的操作写成可复用的 WebMCP 工具。后续任务可以调用这些工具,不必由用户逐步编写网站操作教程。工具的源码和版本保存在本地,可以查看、修改和重新验证。
WebMCP 是什么
WebMCP 是一项提议中的 Web API。网站可以将 JavaScript 功能或 HTML 表单声明为工具,提供名称、自然语言描述和结构化参数。Agent 据此了解可用功能、传入参数并调用工具,与用户共享当前网页和浏览器上下文。
例如,搜索工具的描述说明能搜索什么,参数结构说明需要提供哪些输入。模型可以直接使用这些信息。WebMCP 在网页中提供工具,可以与后端 MCP 集成配合使用。
与普通 Computer Use 的区别
以搜索关键词为例,Computer Use 通常需要找到搜索框、输入、提交、重新读取页面,再提取结果。有对应的 WebMCP 工具时,Agent 可以调用搜索工具并读取返回结果,必要时再调用结果读取工具。
DeepDeck 使用浏览器操作来探索网站和验证工具,后续任务复用已保存的工具。工具未覆盖的操作仍由浏览器完成。
构建工具需要额外的探索和验证,工具描述本身也占用上下文。节省多少时间和 token,取决于网站、任务和工具设计,需要实际测量。
1. 调用网站已有的 WebMCP
打开网站后,DeepDeck 在 Site Agent → WebMCP → Website 下列出发现的工具。切到 Use 并描述任务,Agent 可以调用这些工具。
图中,DeepDeck 在 openai.com 发现了
search_openai。2. 为没有 WebMCP 的网站构建工具
打开 WebMCP Builder,或把 Site Agent 切到 Builder,描述需要的功能。例如:
Agent 在真实页面上尝试操作,检查结果,并将可行的流程写成工具。生成的代码保存在 DeepDeck 中,按网站加载,无需修改目标网站的源码或另行部署 MCP 服务。
以下是 X 上的构建和使用过程。
构建与验证。 Builder 列出生成的工具及验证结果,包括搜索、页面读取和草稿编辑。
查看工具。 这个示例生成了 23 个工具,包括读取账户状态、帖子、个人资料和搜索状态等。工具显示在 Built with DeepDeck 下,与网站提供的工具区分来源。启用后,再次访问会自动加载,也支持查看源码、禁用和回退版本。
调用工具。 切回 Use,输入
compose a hello world x post,Agent 调用 WebMCP 准备Hello, world! 👋草稿。图中没有发布帖子;填写草稿与提交发布是分开的操作。这些工具保存了 Agent 已经验证过的网站操作,供后续任务复用。现有网站因此可以提供 Agent 可调用的功能。网站变化后,相应工具可能需要重新验证和修改。
对 Harness 浏览器插件的建议
社区已有一条 Support webMCP 的讨论。结合 DeepDeck 的实现,我建议浏览器类插件支持以下功能:
DeepDeck 的浏览器界面和 Host 服务通过 Cordis 插件接入,复用 Harness 的会话与 Composer。Electron 提供原生浏览器窗口和页面执行桥接。当前实现需要 DeepDeck 的桌面浏览器环境。
大家有哪些需要反复操作的网站流程?哪些工具发现、构建或版本管理功能,适合由 Harness 的浏览器类插件共用?
DeepDeck 源码 · 图文介绍 · 下载最新版(Apple Silicon / Intel Mac)
English
I built a WebMCP browser in DeepDeck, an unofficial community desktop client based on DeepSeek Harness. It can call tools a website already provides and let an agent build tools for sites without WebMCP.
The user describes a goal. The agent tries the website's controls, checks the results, and writes verified operations as reusable WebMCP tools. Later tasks can call those tools without the user writing step-by-step website instructions. Source code and versions are stored locally for inspection, modification, and verification.
What WebMCP does
WebMCP is a proposed web API. Websites can expose JavaScript functions or HTML forms as tools with names, natural-language descriptions, and structured parameters. Agents use this information to discover capabilities and call tools in the current page, sharing the user's browser context.
For example, a search tool describes what it searches and which inputs it accepts. The model can use that information directly. WebMCP provides tools within a web page and can work alongside backend MCP integrations.
Compared with ordinary Computer Use
A keyword search through Computer Use usually involves finding the search field, typing, submitting, reading the new page, and extracting results. With suitable WebMCP tools, the agent can call a search tool and read its result, calling a separate results-reading tool if needed.
DeepDeck uses browser interaction to explore websites and verify tools. Later tasks reuse saved tools, with browser interaction available for anything they do not cover.
Building tools takes additional exploration and verification, and tool descriptions consume context. Time and token savings depend on the site, task, and tool design and need to be measured.
1. Call existing WebMCP tools
After opening a website, DeepDeck lists discovered tools under Site Agent → WebMCP → Website. Switch to Use and describe the task so the agent can call them.
This screenshot shows
search_openaidiscovered on openai.com.2. Build tools for a site without WebMCP
Open WebMCP Builder, or switch the Site Agent to Builder, and describe the capabilities you need. For example:
The agent tries operations on the live page, checks their results, and implements working workflows as tools. Generated code is saved in DeepDeck and loaded for the matching site. This requires no changes to the target website's source code or a separate MCP server.
The following screenshots show the process on X.
Build and verify. Builder reports the tools it created and the workflows it checked, including search, page reading, and draft editing.
Inspect the tools. This example produces 23 tools, including account-state, post, profile, and search-state readers. They appear under Built with DeepDeck, separately from website-provided tools. Enabled tools load on later visits. You can inspect their source, disable them, or roll back a version.
Call the tools. Switch back to Use and ask
compose a hello world x post. The agent calls WebMCP to prepare aHello, world! 👋draft. The screenshot does not show a published post; filling a draft and submitting it are separate operations.These tools preserve website operations the agent has verified so later tasks can reuse them. Existing websites can gain agent-callable functionality. When a website changes, its tools may need to be checked and updated.
Suggestions for Harness browser plugins
There is an existing Support webMCP discussion. Based on the DeepDeck implementation, I suggest browser plugins support:
DeepDeck's browser UI and Host services integrate through Cordis plugins and reuse Harness sessions and Composer. Electron provides native browser windows and the bridge for page execution. The current implementation requires DeepDeck's desktop browser environment.
Which website workflows do you repeat? Which discovery, tool-building, or version-management features would be useful to share across Harness browser plugins?
DeepDeck source · Illustrated guide · Download the latest release (Apple Silicon / Intel Mac)
All reactions