AI News
商业产业 Hacker News AI 2026-09-30T15:20 正文 ok 918 words

Show HN：我们如何通过构建缺失的记忆层，让 Web Agent 速度提升 3 倍

英文标题：🔶 Show HN: How we made web agents 3x faster by building the missing memory layer

Reduck 推出 Reduck MCP，将可复用脚本与元数据作为浏览器操作的“记忆层”，旨在减少 Agent 每次面对网站时重新探索 UI 的成本。其测试显示，在下载发票、查找 X 热门创始人帖子和根据模糊记忆检索 Reddit 帖子三项任务中，Claude 搭配 Reduck MCP 的整体完成速度提升 2.5 倍，并减少 4.2 倍工具调用。具体而言，下载 5 家服务商的 20 张发票耗时从 472 秒降至 132 秒，即快 3.6 倍。Reduck 还开源了用于 Web Agent 对比评测的 Reduck Voyager，并提供 Claude Code、Codex 等 MCP 客户端接入方式。
打开原文相关 3 · 机会 1
media9.9 KB media404.3 KB media916.1 KB media174.0 KB media151.4 KB media119.4 KB media463.5 KB media196.9 KB media9.2 KB media300.9 KB

中文

可靠且快速地使用浏览器仍是一项挑战

2026 年是编程、网络安全和数学 Agent 的一年，但更广泛的 B2B 工作场景（金融、运营、增长等）以及 B2C 自动化场景（电商、旅行等）尚未受到同等程度的影响。

其中一个关键原因很简单：Agent 根本没有易于使用的 API 来集成这些系统。

如今，即使 Agent 可以利用 Computer Use 自动操作没有 API 的网站（Amazon、Reddit、你的定制 ERP），但对于复杂且高负载的工作流，Computer Use 仍不够胜任，因为它：

  • 缓慢：每次点击都是一个循环（截图、思考、点击、再次截图）。
  • 昂贵：每一步都会将截图和页面快照放入上下文，因此一个任务会消耗大量 token，并触及速率限制。
  • 不稳定：每次运行的结果都有差异，且随着上下文增长会进一步恶化。

Agent 缺少浏览器记忆层

Claude 被要求通过浏览器列出 Notion 发票：它打开 notion.so、读取页面文本、捕获页面、再次打开 notion.so、搜索 Settings、点击并等待；32 秒后仍没有找到发票 Agent 每次都必须手动探索并完成导航

归根结底，Agent 在使用 Computer Use 时既慢又不可靠，原因在于它们：

  • 浪费时间学习网站如何运作：例如，发票需要点击“Settings”而非“Profile”才能找到；还要一直滚动到底部，再点击“Billing”。
  • 浪费时间重复执行操作：即使某项 Skill 保存了导航方式，Agent 仍需在循环中点击和输入、检查新状态、思考，然后再次行动，这会迅速占满其上下文。

推出 Reduck MCP

Reduck 架构：AI Agent 通过 Reduck 的 MCP 和扩展程序访问任何网站

因此，我们构建了 Reduck MCP——用于浏览器操作的记忆层。

Agent 不再需要从头浏览网站或手动执行操作，而是可以编写关于如何使用网站的脚本和元数据，并在未来的导航中更快地复用它们。

为解决冷启动问题，我们维护了一套面向最常用网站的官方脚本库，包括 LinkedIn、Google Maps、Instagram 和 Stripe。

reduck.ai 上的官方脚本库：搜索 LinkedIn 用户、获取 Instagram 个人资料、获取 YC 公司资料、搜索 Google Maps 地点、下载 Stripe 发票 可供 Agent 直接使用的官方脚本

以下是一个“列出 Anthropic Console 发票”脚本的示例。它声明了接受的输入（无）和返回的输出（该组织的所有发票），因此 Agent 可以像调用其他任意工具一样调用它，同时也能知道如何将这些工具串联起来。

“列出 Anthropic Console 发票”脚本页面：脚本功能、无输入，以及包含 count、invoices、organizationId 和 pages 的输出 脚本会说明其功能、接受什么输入及返回什么结果

如果没有匹配的脚本，Agent 可以通过 Reduck MCP 编写一个，并在之后复用。

演示

下面展示 Claude 搭配 Reduck MCP，与 Claude 搭配 Claude for Chrome MCP 下载来自 5 家服务商的 20 张发票的对比：

同一条提示词：左侧是 Claude 搭配 Reduck MCP，右侧是 Claude 搭配 Claude for Chrome MCP。

你可以在 Reduck Voyager 上实时回放这个示例。

那么，Reduck 在这里为什么会快 3.6 倍？

  • 已经了解网站的脚本：脚本目录为这 5 家服务商中的每一家都提供了发票脚本，因此 Agent 调用 run_script，而不是自行摸索各个账单页面。这意味着不再需要进行“读取状态、思考、操作、再次读取”的循环。
  • 并行执行：run_script 接受一批参数，因此脚本可以在多个标签页中并行运行。

在我们运行的每一项任务中都更快

发票只是三项任务中的一项。我们又给 Claude 安排了两项任务，并针对每种工具各使用一次相同提示词：查找本周在 X 上走红的创始人帖子，以及根据模糊记忆查找一篇 Reddit 帖子。使用 Reduck MCP 后，Claude 在全部三项任务中都率先完成，整体速度快 2.5 倍：

每项任务耗时，Claude 搭配 Reduck MCP 对比 Claude 搭配 Chrome MCP：下载 20 张发票为 132 秒对 472 秒（快 3.6 倍）；查找 X 上走红的创始人帖子为 216 秒对 441 秒（快 2.0 倍）；查找一篇 Reddit 帖子为 34 秒对 60 秒（快 1.8 倍） 每项任务耗时：Claude 搭配 Reduck MCP 对比 Claude 搭配 Chrome MCP

时间差异来自无需围绕 UI 反复循环：在 Chrome 中，每次截图、点击和页面读取都是一次工具调用，Claude 会在每次调用前进行思考。使用脚本时，一次调用就能完成一个网站上的整项工作，因此 Claude 在三项任务中的工具调用次数减少了 4.2 倍：

每项任务的工具调用次数，Claude 搭配 Reduck MCP 对比 Claude 搭配 Chrome MCP：下载 20 张发票为 20 次对 132 次（减少 6.6 倍）；查找 X 上走红的创始人帖子为 18 次对 31 次（减少 1.7 倍）；查找一篇 Reddit 帖子为 6 次对 20 次（减少 3.3 倍） 每项任务的工具调用次数：每一次都是 Claude 需要思考的一个步骤

使用 Reduck Voyager 创建你自己的评测

为了生成此类评测及其一对一比较，我们开发了开源 Web Agent 评测框架 Reduck Voyager。更多示例见 voyager.reduck.ai。

它封装了 Claude Code SDK，以完全相同的提示词启动 Agent，唯一的区别是将“Use Reduck MCP”替换为“Use Claude for Chrome MCP”。

评测会尽可能使用确定性代码完成，或者采用人工或 AI 提供的自定义标签。

我们的样本规模较小，欢迎贡献你自己的数据！

几分钟内开始使用

如果你想测试 Reduck MCP，以自动化 Reddit、WhatsApp 或 LinkedIn 上的复杂流程，只需几分钟即可开始：

  1. 在 reduck.ai/#signin 创建账户。
  2. 安装我们的 Chrome 扩展程序，并将其与你的账户配对。
  3. 安装 Reduck MCP。

在 Claude Code CLI 中：

claude mcp add reduck --transport http --scope user https://mcp.reduck.ai

在 Codex CLI 中：

codex mcp add reduck --url https://mcp.reduck.ai

对于其他 MCP 客户端，请参阅我们的接入指南。

启动一个新会话，开始你的首个自动化任务。你可以尝试以下提示词：

Using Reduck MCP, search on Google the top 3 latest posts of the week on "AI Agents" on LinkedIn. Then return the profiles of potential buyers of B2B AI agents.

了解更多

  • 核心概念：关键概念，例如脚本和浏览器执行。
  • API：如何通过 HTTP 端点以编程方式调用 Reduck 脚本，而不是通过 MCP 调用。
  • CLI：适用于并行调用的强大替代方案，可将脚本运行输出保存至本地磁盘，并将其通过管道传递给其他应用。
  • Discord：与我们讨论问题、想法及更多内容。

English / 原文

Reliable and fast browser use is still a challenge

2026 has been the year for coding, cyber and math agents, but broader B2B work (finance, ops, growth, etc.) and B2C automation (e-commerce, travel, etc.) haven’t been as impacted yet.

One of the key reason is simple: Agents simply do not have an easy to use APIs to integrate these systems.

Today, even if agents can use Computer Use to automate sites with no API (Amazon, Reddit, your custom ERP), Computer Use is inadequate for complex and heavy workloads because it’s:

  • Slow : every click is a loop (screenshot, think, click, screenshot again)
  • Expensive : screenshots and page snapshots go into context at every step, so a task burns tokens and hits rate limits
  • Flaky : results vary from run to run and get worse as the context grows

Agents have no memory layer for the browser

Claude, asked to list Notion invoices with its browser, opens notion.so, reads the page text, captures the page, opens notion.so again, searches for Settings, clicks and waits: 32 seconds in, still no invoice Agents have to manually discover and navigate each time

At the heart of it, agents are slow and unreliable at Computer Use because they:

  • Waste time learning how a website works : e.g. that invoices are found by clicking “Settings” and not on “Profile”, scrolling all the way down, and clicking “Billing”.
  • Waste time redoing the action : even with a Skill that stores how to navigate, the agent still operates in a loop to click and type, check new state, think and act again, which floods its context.

Introducing Reduck MCP

Reduck architecture: AI agents connect through Reduck's MCP and extension to reach any website

That’s why we built Reduck MCP, the memory layer for browser actions.

Instead of browsing from scratch or performing manual actions, Agents can now write scripts + metadata on how to use it and then use it for faster future navigation.

To solve the cold start problem we maintain a library of official scripts for the most used sites, such as LinkedIn, Google Maps, Instagram and Stripe.

The official scripts library on reduck.ai: search LinkedIn people, get an Instagram profile, get a YC company profile, search Google Maps places, download a Stripe invoice The official scripts, ready for your agent

Here is an example of script with “List Anthropic Console invoices”. It states the input it takes (none) and the output it returns (every invoice of the organization), so an agent calls it the way it calls any other tool, but also can know how to chain them.

The page of the List Anthropic Console invoices script: what it does, no input, and an output of count, invoices, organizationId and pages A script states what it does, what it takes and what it returns

If no script matches, agents can write one with Reduck MCP and use it later.

Demo

Here is Claude with Reduck MCP, against Claude with the Claude for Chrome MCP, downloading 20 invoices from 5 providers:

Same prompt: Claude with Reduck MCP on the left, Claude with the Claude for Chrome MCP on the right

You can replay this example live on Reduck Voyager.

So how is Reduck 3.6x faster here?

  • Scripts that already know the site : the catalogue has an invoice script for each of the five providers, so the agent calls run_script instead of finding its way around each billing page. This means that there is no need to do the loop with read state, think, atc, read again.
  • Parallel execution : run_script takes a batch of args, so scripts run in parallel on several tabs.

Faster on every task we ran

The invoices are one task of three. We gave Claude two more, each with the same prompt once per tool: find this week’s viral founder posts on X, and find a Reddit post from a vague memory. With Reduck MCP, Claude finished all three first, 2.5x faster overall:

Time per task, Claude with Reduck MCP against Claude with Chrome MCP: 132 s against 472 s to download 20 invoices (3.6x faster), 216 s against 441 s to find viral founder posts on X (2.0x faster), 34 s against 60 s to find a Reddit post (1.8x faster) Time per task: Claude with Reduck MCP against Claude with Chrome MCP

The time difference comes from not having to loop on the UI: with Chrome, every screenshot, click and page read is a tool call, and Claude thinks before each one. With a script, one call does the whole job on a site, so Claude made 4.2x fewer tool calls over the three tasks:

Tool calls per task, Claude with Reduck MCP against Claude with Chrome MCP: 20 against 132 to download 20 invoices (6.6x fewer), 18 against 31 to find viral founder posts on X (1.7x fewer), 6 against 20 to find a Reddit post (3.3x fewer) Tool calls per task: each one is a step Claude has to think about

Create your own evals with Reduck Voyager

To produce such evaluations and their 1:1 comparison, we wrote Reduck Voyager, our open-source eval framework for web agents. Other examples are at voyager.reduck.ai.

It wraps the Claude Code SDK to launch agents on exactly the same prompts, except that “Use Reduck MCP” becomes “Use Claude for Chrome MCP”.

Evaluation is done with deterministic code when possible, or with custom labels from a human or an AI.

Our sample is small, so feel free to contribute your own data!

Get started in minutes

If you want to test Reduck MCP to automate complex flows on Reddit, WhatsApp, or LinkedIn, you can get started in literally minutes:

  1. Create an account at reduck.ai/#signin
  2. Install our Chrome extension and pair it with your account
  3. Install Reduck MCP

On Claude Code CLI:

claude mcp add reduck --transport http --scope user https://mcp.reduck.ai

On Codex CLI:

codex mcp add reduck --url https://mcp.reduck.ai

For other MCP clients, see our Onboarding Guide.

Start a new session and get going with your first automation. You can try a prompt like:

Using Reduck MCP, search on Google the top 3 latest posts of the week on "AI Agents" on LinkedIn. Then return the profiles of potential buyers of B2B AI agents.

Learn more

  • Core concepts: key concepts, such as scripts and browser execution
  • API: how to call Reduck scripts programmatically from HTTP endpoints instead of MCP calls
  • CLI: a strong alternative for parallel calls, saving script run outputs to disk locally, and piping them to other apps
  • Discord: chat with us about issues, ideas, and more