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。

发布出去的产品只有两样：命令行 envoy，和 Claude Code / Codex / DeepSeek Harness 的插件与 skills，都装在本机。仓库根目录 docker-compose.yml 那套 Web 界面（FastAPI + Next.js + MySQL）是遗留选项，装了 CLI 也不需要它。Python 包目前只随 GitHub Release 发 envoy_cli-<版本>-py3-none-any.whl，PyPI 上还没有发布。

60 秒演示：一条真实链接 → 直接拿到 [mm:ss] 字幕片段 → 没读到的范围显式列为待覆盖（读 TED 演讲 0–30 秒）。

录制与复现命令见 docs/demo/README.md。

各站点实际能读到什么（YouTube / B 站 / Vimeo / Dailymotion / Instagram 的真实实测，含网络不通时自动走可信代理）见 站点可读性说明。

三个宿主各一行。安装脚本只要求 curl：uv 和 Python 3.11 它自己装，装好会打印 envoy install: runtime … protocol … browser …。

# Claude Code
curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- claude-code

# Codex
curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- codex

# DeepSeek Harness
curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- dsh

默认是最小档：只装 runtime（≈178 MB，读普通网页、PDF、图片），Chromium 与 Whisper 权重都不在安装包里。第一次读需要渲染的页面时，错误直接给出 envoy browser setup（一次性下载 Chromium，约 683 MB，1–2 分钟）；第一次需要转写（字幕、语音、无字幕视频）时给出 envoy asr install --model base（权重约 140 MB）。不想等到首次使用：--profile standard（runtime + Chromium，≈860 MB）或 --profile full（再加 Whisper base，装完全部能力离线可用）。重跑安装脚本并换成更高的档位就能补齐缺的能力，两步都是幂等的；体积口径与平台差异见 INSTALL.md。

不确定宿主是哪个：把最后一段换成 auto，脚本会识别已装的 Claude Code / Codex / DSH，都没有就写 MCP 条目（mcp、skill 也可以单独装；all 等同于 auto）。只想先看它要做什么，命令前加 ENVOY_DRY_RUN=1。从镜像或本地目录安装时，把下载根换成 ENVOY_RELEASES_BASE、清单地址换成 ENVOY_MANIFEST_URL（目录按 v<版本>/ 放产物与 SHA256SUMS），默认值不变；整仓换到 fork 时设 ENVOY_REPO（默认 regrevia/Envoy）。Windows PowerShell 用 irm https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.ps1 | iex（默认 auto，指定宿主时下载后 powershell -ExecutionPolicy Bypass -File install.ps1 dsh）。

装完两条自检：

envoy doctor --json     # runtime.install.status 应为 pinned；最小档下 browser.ready 与 asr_model_available 先为 false，补齐后应为 true
envoy browser setup     # 最小档装完没有 Chromium；这行补装（约 683 MB），也用于换版本，装完会自动清掉被替换的旧版本
envoy asr install --model base   # 最小档没有 Whisper 权重；这行补装（约 140 MB）
envoy browser prune     # 回收旧版本浏览器下载（升级几次后缓存会翻倍），--dry-run 只报不删

升级就是重跑同一行；卸载、企业镜像、各宿主插件目录见 INSTALL.md。

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。

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]

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）。

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。

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，题目与来源冻结在 benchmarks/fixed-set-v1。

百分比是各臂组内占比。两臂时间上限不同（900 秒 vs 2400 秒），完成率差距会被放大；verified 两臂都是 0，即本轮没有人工判定答案对错，suspect_complete 只表示「正常结束且没说读不到」。

判定口径、偏差与复核记录见评测方法、口径与偏差与复核记录。

123 条真实链接（新闻、公众号、B站、播客、PDF、论坛、代码文档……）跑了一遍，看三件事：

「可用」= 能当素材：整篇可引用，或读到一部分并写明缺了什么；读不出来会说清原因（要登录、被人机校验拦住、网络不通），不猜。

剩下没拿到的 24 条里，16 条根本连不上（全是境外站点，卡在本机代理出口），8 条卡在登录墙或人机校验——这两种在本轮实验的条件下装不装 Envoy 都读不到，Envoy 会直说，让 Agent 走别的路。

装不装 Envoy 的对照见评测。复算：scripts/acceptance/link_coverage.py（报告与口径见兼容性实验）。

插件与 skills 不依赖 Docker。 Claude Code / Codex / DSH 三个装法都是本机跑：用 install.sh / install.ps1 安装脚本，或从 GitHub Release 下 envoy_cli-<版本>-py3-none-any.whl 交给 uv tool install，再加一个 skills 目录或插件包（PyPI 上还没有发布包），不需要容器、不需要这个 compose。Docker 只服务仓库里那套旧的 Web 界面（FastAPI + Next.js + MySQL），用来自本机点着看任务和产物；不用这个界面就完全不用管它。

需要 Docker Desktop，仓库根目录：

cp .env.example .env
docker compose up -d --build

前端 http://127.0.0.1:8200，后端健康检查 http://127.0.0.1:8140/api/v1/health 返回 {"status":"ok"}。docker compose ps 里 mysql、redis、backend、frontend 都 healthy 即可用；MySQL 首次启动会自动执行 sql/ 下的建表脚本，上传与导出都写在 envoy_runtime 卷里，默认不需要对象存储。首次构建会拉 MySQL 8.4、Redis，并在后端镜像里装 FFmpeg、Tesseract、Chromium，耗时几分钟。docker compose down 保留数据卷；只有 down -v 会删库和本地文件。端口、密码、模型配置都在 .env，构建加速与排错见 Docker 一键启动。

• 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
• 链接覆盖基线与复测
• 登录扫码交接
• 版本记录
• 安全与隐私说明

中文

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.

One line per host. The installer only needs curl: it installs uv and Python 3.11 itself and prints envoy install: runtime … protocol … browser … when done.

# Claude Code
curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- claude-code

# Codex
curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- codex

# DeepSeek Harness
curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- dsh

The default is the minimal profile: the runtime alone (about 178 MB - plain pages, PDFs, images), with Chromium and the Whisper weights left out of the install. The first page that needs rendering reports the exact command that fetches Chromium (envoy browser setup, a one-time ~683 MB download, 1-2 minutes), and the first transcript reports envoy asr install --model base (~140 MB of weights). A box that must not download later passes --profile standard (runtime + Chromium, about 860 MB) or --profile full (adds Whisper base, so every capability works offline after install). Re-running the installer with a higher profile fills in the rest; both steps are idempotent, and INSTALL.md has the size accounting and platform differences.

Not sure which host you have: use auto instead — it detects an installed Claude Code / Codex / DSH and falls back to an MCP entry (mcp and skill also install on their own; all is the same as auto). Add ENVOY_DRY_RUN=1 to see what it would do first. To install from a mirror or a local directory, point ENVOY_RELEASES_BASE at the download root and ENVOY_MANIFEST_URL at the manifest (layout: v<version>/ with the artifacts and SHA256SUMS); the defaults are unchanged; to point at a fork set ENVOY_REPO (default regrevia/Envoy). On Windows PowerShell: irm https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.ps1 | iex (auto by default; for one host download it and run powershell -ExecutionPolicy Bypass -File install.ps1 dsh).

Two checks after installing:

envoy doctor --json     # runtime.install.status should be pinned; the minimal profile starts with browser.ready and asr_model_available false, and both turn true once fetched
envoy browser setup     # the minimal install has no Chromium; this fetches it (~683 MB), and repairs or changes versions, pruned stale copies included
envoy asr install --model base   # the minimal install has no Whisper weights; this fetches them (~140 MB)
envoy browser prune     # reclaim browser downloads left behind by upgrades; --dry-run only reports

Upgrading means running the same line again. Uninstall, mirrors, and each host's plugin directory: INSTALL.md.

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.

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]

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).

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.

Same link, same model, the only difference is whether Envoy is installed: 20 task packages × 3 questions per arm (60 turns each) on a frozen package (benchmarks/fixed-set-v1).

Percentages are within-arm shares. The arms ran under different caps (900 s vs 2400 s), which inflates the completion gap; verified is 0 for both, meaning no human correctness review, and suspect_complete only means "exited normally without saying it could not read".

Scoring definitions, deviations and the review record are in the evaluation method, the protocol and deviations and the review record.

123 real links (news, WeChat posts, Bilibili, podcasts, PDFs, forums, code docs) — three questions:

"Usable" means quotable in full, or partially read with the gap named; a failure states why (login, bot check, network) instead of guessing.

Of the 24 that were not usable, 16 hosts were unreachable — all overseas, blocked at the local proxy exit — and 8 sat behind a login or bot-check wall. Under the conditions of this run neither case reads with or without Envoy installed; Envoy says so and the agent takes another route.

With vs without Envoy: see Evaluation. Rerun with scripts/acceptance/link_coverage.py; reports and definitions are in the link-compatibility experiment.

The plugin and skills do not need Docker. All three installs (Claude Code, Codex, DSH) run locally: use the install.sh / install.ps1 installer, or download envoy_cli-<版本>-py3-none-any.whl from the GitHub Release and hand it to uv tool install, plus a skills directory or plugin package (PyPI has no published package yet) — no container, no compose. Docker only serves the legacy web interface in this repo (FastAPI + Next.js + MySQL) for browsing tasks and artifacts on your own machine; skip it entirely if you do not want that UI.

With Docker Desktop running, from the repository root:

cp .env.example .env
docker compose up -d --build

The interface is at http://127.0.0.1:8200; http://127.0.0.1:8140/api/v1/health answers {"status":"ok"}. The stack is ready when mysql, redis, backend and frontend all report healthy; MySQL bootstraps the schema from sql/ on first start, and uploads plus exports stay on the envoy_runtime volume, so no object storage is required. The first build pulls MySQL 8.4 and Redis and installs FFmpeg, Tesseract and Chromium into the backend image, so it takes a few minutes. docker compose down keeps the volumes; only down -v drops the database and local files. Ports, passwords and model settings live in .env; build acceleration and troubleshooting are in Docker quickstart.

• Installation, updates, and MCP setup
• Claude Code plugin
• Codex plugin
• DeepSeek Harness plugin
• Product and engineering status
• Platform acceptance matrix
• Link coverage baseline and re-run
• Sign-in QR handoff
• Changelog
• Security and privacy