Envoy (envoy-dsh)

把链接分享给你的 AI

一个链接，成为你想要的任何东西。

安装 Envoy · 下载发行版 · 反馈问题 · English

给 AI 一个链接，Envoy 把它变成带来源定位的材料，AI 再用它写你要的东西：Wiki、摘要、报告、脚本。网页、公众号、PDF、小红书图文、B 站视频、本地视频都行。

能做什么

关键帧 — 视频里的 PPT、图表、演示画面按时间点抽帧交给 AI 看，不用来回拖进度条。

视频理解 — 优先读公开字幕；没有字幕时对指定时间段跑本地 Whisper，返回带 [mm:ss] 的转写片段。

图文 — 网页、公众号文章、PDF、图片帖，变成干净正文、图片和来源。

知识库 / Wiki — 长视频整段读完，按主题分节，每条结论带时间戳，直接放进 Obsidian、Notion。

怎么用

对 AI 说：

把这条视频整理成知识库：https://www.bilibili.com/video/BV...

先读前 10 分钟，按主题分节；每条结论后标注 [mm:ss]；没覆盖到的部分单列“待补”，不要猜。

材料到手就已经分好段、带定位，所以产出可以直接引用：

## 2. 为什么选择固定轮

- 固定轮把连续动作拆成可复用的阶段，便于逐段调试。[04:18]
- 作者对比了三种方案，并给出取舍条件。[07:02]
- 这一段没有展示实际测试数据；待补。[09:31]

工作原理

1. 取材料 — 打开链接，拿到标题、正文、字幕或转写、PDF 文本、图片。
2. 带定位 — 切成片段并附 video_time 或来源坐标，给一个 material id。
3. 按需续读 — envoy_search 定位关键词，envoy_expand 补齐没读的部分，不重拉整页。
4. 缺口照实报 — 没字幕、要登录、模态缺失、下载预算不够，标记为未覆盖，不写成结论。

浏览器在后台运行，遇到登录或验证才请你接管。接入方式是 Agent Skill + MCP，Claude Code、Codex、DeepSeek Harness 各有插件（Codex 0.156 起需执行一次 envoy hook install --host codex）。

边界

不绕过登录、验证码、地区限制和 DRM；不替代联网搜索。Envoy 只对实际读取并返回的时间段负责。抖音主页集合、小红书图文等能力按平台验收矩阵逐项验证。

评测

同一条链接、同一个模型，只差装不装 Envoy；20 个任务包 × 3 轮提问（每臂 60 轮，共 120 轮），题目与来源冻结在 benchmarks/fixed-set-v1。下面是探索性观察，还不足以当作结论。

怎么判的。 每轮先按答案文本归类：timed_out（撞墙钟上限）、no_output、infra_failed、needs_action、cannot_see（要登录/扫码/授权，或说读不到）、suspect_complete（进程正常结束、答案不短、没说读不到——只是疑似完成，不验证事实是否正确、引用是否落位）。verified 只统计人工评审把 answer_correct 判为 true 的轮次；本轮 verified = 0，评审表未填。facts 命中是 60% 词覆盖匹配，引用命中要求 URL 规范化到文章级（同一公众号的不同文章不再互相计分），两者都不是正确性判定。

本轮可复算的观察。 baseline 60 轮：suspect_complete 13、timed_out 41、cannot_see 5、needs_action 1。envoy 60 轮：suspect_complete 55、timed_out 0、cannot_see 4、needs_action 1。两个 verified 都是 0。facts 词覆盖也偏向 envoy（claim 覆盖 9.1% vs 38.2%，至少命中一条 facts 的轮次 14/60 vs 49/60），但同样受下面这条限制影响，只能算线索。

不能这么读。 两臂时间上限不同：baseline 900 秒（/tmp/AB3/run-meta.json 的 cap_seconds；60 轮里 41 轮撞上限），envoy 2400 秒（/tmp/AB6 各轮 turn_timeout_seconds；0 轮撞上限）。完成率差距被上限差异放大，不能读成工具能力差；suspect_complete 也不能读成“答案正确”。要变成结论还缺：同预算比较（两臂同上限重跑）、可复核的真实任务判定（不是关键词分类）、能与来源快照对上的产物，以及两臂完整的 60/60 轮账本。旧版首页给出的完成率与样本数字来自更早的部分运行，本轮无法复现，已撤回。

见评测方法、口径与偏差与复核记录。

文档

• 安装、更新与 MCP 配置
• Claude Code 插件
• Codex 插件
• DeepSeek Harness 插件
• 产品与技术状态
• 平台真实验收矩阵
• 版本记录
• 安全与隐私说明

English

中文

Share a link with your AI

One link in. Anything you want out.

Give your AI a link and Envoy turns it into material with source locators; the AI then writes whatever you asked for: a wiki, a summary, a report, a script. Web pages, WeChat posts, PDFs, Xiaohongshu image posts, Bilibili videos, local video files.

What it does

Keyframes — pull slides, charts, and demo frames at chosen timestamps and hand them to the AI, no scrubbing.

Video — captions first; when there are none, run local Whisper over the requested range and return transcript segments with [mm:ss].

Text and images — web pages, articles, PDFs, image posts become clean text, images, and their source.

Knowledge base / wiki — read a long video end to end, group it by theme, put a timestamp on every claim, and drop the result into Obsidian or Notion.

How to use it

Ask your AI:

Turn this video into a knowledge base: https://www.bilibili.com/video/BV...

Read the first 10 minutes, group sections by theme, and add [mm:ss] after each claim. List anything not covered under "To fill"; do not guess.

The material arrives already segmented and located, so the output can cite directly:

## 2. Why use a fixed wheel

- A fixed wheel splits a continuous action into reusable stages for per-stage debugging. [04:18]
- The author compares three approaches and states the trade-offs. [07:02]
- This section shows no measured data; to fill. [09:31]

How it works

1. Fetch — open the link and take the title, body, captions or transcript, PDF text, and images.
2. Locate — split it into segments carrying video_time or source coordinates, and return a material id.
3. Continue on demand — envoy_search finds a term, envoy_expand fills the unread range; no second full fetch.
4. Report gaps — no captions, login required, missing modality, exhausted download budget: marked as uncovered rather than written up as a conclusion.

The browser runs headlessly and asks you to take over only for login or verification. Access is an Agent Skill plus MCP server; Claude Code, Codex, and DeepSeek Harness each have a plugin (Codex 0.156+ needs one envoy hook install --host codex).

Scope

No bypassing login, CAPTCHA, regional restrictions, or DRM, and no replacement for web search. Envoy is responsible only for the ranges it actually read and returned. Douyin creator collections and Xiaohongshu image posts follow the platform acceptance matrix.

Evaluation

Same link, same model, the only difference is whether Envoy is installed; 20 task packages × 3 questions per arm (60 turns each, 120 total) on a frozen package (benchmarks/fixed-set-v1). What follows is an exploratory observation, not yet a conclusion.

How turns are scored. Each turn is classified from its answer text: timed_out (hit the wall-clock cap), no_output, infra_failed, needs_action, cannot_see (asks for login/scan/authorization, or says it cannot read), suspect_complete (process exited normally, answer is not short, no refusal — suspected only, with no check of factual correctness or citation placement). verified counts only turns a human review marked answer_correct = true; verified = 0 here because the review file is still empty. Fact hits are a 60% token-overlap match, and citation hits require an article-level canonical URL (two posts from the same WeChat account no longer credit each other); neither is a correctness verdict.

What can be recomputed now. baseline, 60 turns: suspect_complete 13, timed_out 41, cannot_see 5, needs_action 1. envoy, 60 turns: suspect_complete 55, timed_out 0, cannot_see 4, needs_action 1. verified is 0 for both. Fact-word coverage also leans envoy (claim coverage 9.1% vs 38.2%; turns with at least one fact 14/60 vs 49/60), but it is subject to the limit below, so treat it as a signal only.

What it cannot be read as. The two arms ran under different caps: baseline 900 s (cap_seconds in /tmp/AB3/run-meta.json; 41 of 60 turns hit it) and envoy 2400 s (turn_timeout_seconds in /tmp/AB6; 0 of 60 hit it). The completion gap is inflated by that difference and is not a capability comparison; suspect_complete is not "correct". Turning this into a conclusion needs a same-budget run (both arms at one cap), a real task-level verdict that can be reviewed (not keyword rules), artifacts that can be checked against the source snapshots, and a full 60/60-turn ledger per arm. The earlier front-page completion rates and sample counts came from an earlier partial run, cannot be reproduced from the archived data here, and have been withdrawn.

See the evaluation method, the protocol and deviations, and the review record.

Documentation

• Installation, updates, and MCP setup
• Claude Code plugin
• Codex plugin
• DeepSeek Harness plugin
• Product and engineering status
• Platform acceptance matrix
• Changelog
• Security and privacy