让每一次截图与录音，都成为可追溯、可复用的长期资产。 | Turn every screenshot and recording into traceable, reusable long-term assets.

---

Language / 语言：中文 | English

---

中文

介绍

AuraCap 是首个将 GitHub Release 作为瞬态中间件（Transient Middleware）的开源项目：iOS 快捷指令把截图或录音上传至 Release Assets，GitHub Actions 自动拉取、AI 处理并写入时间线——全程在 GitHub 安全边界内闭环，无需 S3、无需图床、无需暴露 Webhook。Fork 仓库即完成部署，运行成本为零。

解决什么问题

多数信息系统的记录与整理是两件事：当下来不及整理，事后很难回到现场。而大多数"零成本"方案的代价是暗中依赖第三方存储，或将你锁在特定平台。AuraCap 用一种"黑客精神"解开这个结：让 GitHub 自身的基础设施承担全部中转与处理，让"捕捉"天然通向"沉淀"，数据始终由你掌控。

为什么选择 Shortcuts（快捷指令）作为入口

选择 Shortcuts，不是为了"更酷"，而是为了降低长期使用门槛。

• 贴近真实场景：在通勤、会议、阅读或对话中，需要的是一步可达，而不是切换到重应用再组织流程
• 交互轻量且稳定：快捷指令是系统级能力，调用链路短、交互负担低，适合高频、低摩擦的日常记录
• 系统边界清晰：前端交互极简，后端能力可独立演进；可以替换模型、切换部署方式，而不必重做采集入口
• 更容易被坚持：真正产生价值的记录系统，不是功能最多，而是愿意每天都用

特点

• 入口轻，能力重：采集通过快捷指令完成，整理由 AI 自动执行
• 部署路径灵活：可自部署后端，也可仅依赖 GitHub 工作流；后者无需自建服务器
• 输出长期可控：统一写入 Markdown（如 storage/timeline.md），便于检索、归档、迁移与备份
• 模型生态开放：支持 OpenAI、Gemini、SiliconFlow、Anthropic、Groq、Mistral 等模型及平台

架构设计：为什么选择 GitHub Release 作为"瞬态媒体中转站"

AuraCap 可能是首个提出并实现 "GitHub Release as a Transient Middleware"（将 GitHub Release 作为瞬态中间件）的开源项目。

在设计 AuraCap 时，我面临一个核心矛盾：如何在不依赖付费 S3 存储、不自建图床、且不暴露 Webhook 代理的情况下，让 iOS 快捷指令与 GitHub Actions 传递大体积媒体文件，实现零外部依赖的媒体采集与处理，把 GitHub Release 变成模型的预处理器？

AuraCap 最终采取了一种些许"黑客精神"的方案：利用 GitHub Release Assets 作为异步中转。

1. 打破"静态发布"的刻板印象：通常 Release 用于托管稳定的二进制包，但我将其视为一个 带身份认证的临时对象存储（Object Storage）。
2. 利用 asset_id 建立信任链：
   • iOS 端通过 API 将截图/录音上传至特定的 Inbox Release。
   • 上传成功后，GitHub 返回一个唯一的 asset_id。
   • 快捷指令仅需携带这个 asset_id 触发 workflow_dispatch，而无需传输整个文件。
3. 零配置的安全内循环：
   • GitHub Actions 环境中自带 GITHUB_TOKEN，拥有拉取本仓库 Release Assets 的天然权限。
   • 处理脚本根据 asset_id 直接下载、处理、写入。
   • 闭环完成：整个过程不产生额外的存储费用，不依赖第三方 API Key，且媒体流始终在 GitHub 安全边界内流动。

选择使用方式

方式	条件	特点
GitHub-only	不希望维护服务器，希望尽量简化运维	上传截图至 Release 后触发 Workflow 自动处理
自部署	有服务器，或本机可运行 Python / Docker	iOS 直接连接你的 API，响应快、可控性高

两种方式都可独立使用：前者强调低维护成本与开箱即用，后者强调实时性与控制力。

快速上手

GitHub-only（推荐）

1. Fork 本仓库（隐私提醒：Fork 后请将仓库设为 Private，否则截图/录音会对外可见）
2. 在仓库 Settings -> Secrets and variables -> Actions 中配置必要变量（如 TEXT_PROVIDER、MM_PROVIDER、OUTPUT_LOCALE 等，可先使用 mock 跑通流程；mock 模式无需 Secrets；使用真实模型时需在 Secrets 页签添加 API Key（如 OPENAI_API_KEY）；详见 GITHUB_RELEASE_INBOX 步骤 2）
3. 运行一次 AuraCap Setup Release Inbox 工作流
4. 按用户手册在 iPhone 上搭好快捷指令（或 下载已脱敏模板（截图）、快捷指令模板（录音）、快捷指令模板（拍照），导入后仅填写个人信息即可）

建议首次先使用 mock 模式完成端到端验证，再切换为 OpenAI、Gemini 等真实模型。详细步骤见 用户手册。

自部署（3 步）

git clone <your-fork-or-repo-url>
cd AuraCap
cp .env.example .env
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt
python backend/main.py

或使用 Docker：docker compose up -d --build。浏览器访问 http://127.0.0.1:8000/health，出现响应即表示服务启动成功。首次运行快捷指令时会提示输入 AuraCap Backend Base URL（如 http://192.168.1.23:8000 或 https://cap.yourdomain.com）。

会得到什么输出

• storage/timeline.md：按时间顺序记录条目，包含 AI 提取结果
• storage/insights/：每日洞察（默认每天一次，可配置）
• storage/summary/：定期摘要（默认每周一次，可配置）
• storage/task_index/：任务索引（按 tag 权重的日/周索引，默认每日生成）
• storage/customized/：自定义操作输出（需启用 ENABLE_CUSTOM_OPERATION）

所有结果均为 Markdown，可同步到 Notion、Obsidian 或任意知识管理系统。你的数据结构不会被平台锁定，也便于长期积累与二次利用。调度频率与 cron 配置见 用户手册 3.5 自动化调度。

提示词说明

AuraCap 使用五类提示词驱动 AI 分析，均位于 prompts/ 目录：

提示词	作用	触发时机
timeline_prompts.md	从截图或录音中提取核心信息，写入 timeline	每次截图/录音时
insights_prompts.md	通读当日 timeline，发现跨条目的模式与未完成信号	每日定时（默认 UTC 01:00）
summary_prompts.md	纵向分析一段时间内的 timeline + insights，归纳轨迹与建议	每周定时（默认周日 UTC 02:00）
tagging_zh.md / tagging_en.md	为 timeline 条目生成 1–3 个语义 tag，供任务索引使用	每日定时（与 insights 同 cron）
customized_prompts.md	对 timeline 提取结果做额外 AI 处理，输出到 storage/customized/	每次捕捉后或按 cron 定时（需启用 ENABLE_CUSTOM_OPERATION）

Timeline 支持 4 套场景化提示词（截图/录音 × 中文/英文），insights 与 summary 各支持中英文 2 套；通过 TIMELINE_LANG_MODE 与 OUTPUT_LOCALE 控制语言路由。默认 timeline 提示词针对 iOS 截图 优化（过滤状态栏等系统噪音）。若你主要使用录音，可自行修改 prompts/timeline_prompts.md，或通过 TIMELINE_PROMPT_FILE 指定自己的文件；自定义操作提示词可通过 CUSTOMIZED_PROMPT_FILE 指定路径。OpenAI 兼容：OPENAI_* 变量适用于 OpenAI 官方及 SiliconFlow、OpenRouter、DeepSeek 等兼容服务，接入第三方时改 OPENAI_BASE_URL 即可。详见 用户手册 3.6 提示词。

下一步

• 完整配置与模型选择：docs/USERGUIDE.md
• GitHub-only 详细说明：docs/GITHUB_RELEASE_INBOX.md
• 快捷指令模板与手动搭建：shortcuts/README.md
• 从上游同步时保留 storage：docs/STORAGE_MIGRATION.md

协议

MIT License

---

English

Introduction

AuraCap is the first open-source project to implement GitHub Release as Transient Middleware: iOS Shortcuts upload screenshots or recordings to Release Assets; GitHub Actions then fetches, processes via AI, and writes to a timeline — fully closed-loop within GitHub's security boundary, no S3, no image hosting, no exposed webhooks. Forking the repository is the entire deployment. Running cost: zero.

The Problem It Solves

Most information tools split recording and organizing into two separate acts: you never have time to organize in the moment, and it's hard to return to the context later. Most "zero-cost" solutions quietly depend on third-party storage or lock your data into a platform. AuraCap resolves this with a touch of hacker ingenuity: GitHub's own infrastructure handles all the relay and processing, so "capture" naturally leads to "consolidation" — and you always own your data.

Why Shortcuts as the Entry Point

Choosing Shortcuts is not about being "cooler"—it's about lowering the barrier for long-term use.

• Close to real scenarios: When commuting, in meetings, reading, or in conversations, you need one-tap access, not switching to a heavy app and orchestrating a flow
• Lightweight, stable interaction: Shortcuts are a system-level capability; the call path is short, the interaction burden is low, suitable for high-frequency, low-friction daily capture
• Clear system boundaries: Minimal frontend, backend can evolve independently; you can swap models and deployment modes without rebuilding the capture entry
• Easier to stick with: A recording system that actually adds value is not the one with the most features, but the one you're willing to use every day

Features

• Light entry, heavy capability: Capture via Shortcuts, organization by AI
• Flexible deployment: Self-host a backend or rely solely on GitHub workflows; the latter requires no server
• Long-term control over output: Everything written to Markdown (e.g. storage/timeline.md), easy to search, archive, migrate, and back up
• Open model ecosystem: Supports OpenAI, Gemini, SiliconFlow, Anthropic, Groq, Mistral, and other models/platforms

Architecture: Why Use GitHub Release as a "Transient Media Relay"

AuraCap may be the first open-source project to propose and implement "GitHub Release as a Transient Middleware."

The core design challenge was: How can iOS Shortcuts and GitHub Actions pass large media files without paid S3 storage, self-hosted image hosting, or exposed webhook proxies—achieving zero external dependencies for media capture and processing, turning GitHub Release into the model's preprocessor?

AuraCap's answer: Use GitHub Release Assets as an asynchronous relay.

1. Rethinking Release: Instead of treating Release only as a place for stable binaries, we treat it as authenticated temporary object storage.
2. Trust chain via asset_id:
   • iOS uploads screenshots/recordings to a dedicated Inbox Release via API.
   • GitHub returns a unique asset_id.
   • The shortcut passes only this asset_id to trigger workflow_dispatch—no need to transmit the full file.
3. Zero-config secure loop:
   • GitHub Actions has GITHUB_TOKEN with built-in permission to fetch the repo's Release Assets.
   • The processing script downloads, processes, and writes based on asset_id.
   • Closed loop: No extra storage cost, no third-party API keys, and the media flow stays within GitHub's security boundary.

Choose Your Mode

Mode	Requirement	Characteristics
GitHub-only	No server maintenance; minimal ops	Upload to Release, trigger Workflow; processing is automatic
Self-host	Server or local Python/Docker	iOS connects directly to your API; fast response, full control

Both can be used independently: the former emphasizes low maintenance and out-of-the-box use, the latter emphasizes real-time control.

Quick Start

GitHub-only (recommended)

1. Fork this repository (Privacy: Set your fork to Private, or screenshots/recordings will be publicly visible.)
2. Configure required variables under Settings -> Secrets and variables -> Actions (e.g. TEXT_PROVIDER, MM_PROVIDER, OUTPUT_LOCALE; use mock first to verify; mock mode needs no Secrets; for real models add API key in Secrets (e.g. OPENAI_API_KEY); see GITHUB_RELEASE_INBOX Step 2)
3. Run the AuraCap Setup Release Inbox workflow once
4. Set up the shortcut on your iPhone following the user guide (or Screenshot template, Voice template, Photo template—fill in your personal info only)

Start with mock mode to complete an end-to-end run, then switch to OpenAI, Gemini, or other real models. See docs/USERGUIDE.md for details.

Self-host (3 steps)

git clone <your-fork-or-repo-url>
cd AuraCap
cp .env.example .env
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt
python backend/main.py

Or use Docker: docker compose up -d --build. Visit http://127.0.0.1:8000/health in your browser; a response means the service is up. On first run, the shortcut will prompt for AuraCap Backend Base URL (e.g. http://192.168.1.23:8000 or https://cap.yourdomain.com).

Output

• storage/timeline.md: Time-ordered entries with AI-extracted content
• storage/insights/: Daily insights (default: once per day, configurable)
• storage/summary/: Periodic summaries (default: once per week, configurable)
• storage/task_index/: Task index (tag-weighted daily/weekly index, generated daily by default)
• storage/customized/: Custom operation output (requires ENABLE_CUSTOM_OPERATION)

All output is Markdown and can be synced to Notion, Obsidian, or any knowledge management system. Your data format stays platform-agnostic and supports long-term accumulation and reuse. For schedule frequency and cron configuration, see User Guide 3.5 Scheduler.

Prompts

AuraCap uses five prompt files under prompts/ to drive AI analysis:

Prompt	Purpose	Trigger
timeline_prompts.md	Extract key info from screenshots or recordings into timeline	On each capture
insights_prompts.md	Analyze the day's timeline for patterns and open threads	Daily (default UTC 01:00)
summary_prompts.md	Longitudinal analysis of timeline + insights over a period	Weekly (default Sunday UTC 02:00)
tagging_zh.md / tagging_en.md	Assign 1–3 semantic tags per timeline entry for task indexing	Daily (same cron as insights)
customized_prompts.md	Extra AI processing on timeline extract results, output to storage/customized/	After each capture or on cron schedule (requires ENABLE_CUSTOM_OPERATION)

Timeline supports four scenario-specific prompts (screenshot/audio × Chinese/English); insights and summary each support two language variants. Language routing is controlled by TIMELINE_LANG_MODE and OUTPUT_LOCALE. The default timeline prompt is tuned for iOS screenshots (filtering status bar etc.). If you mainly use voice recordings, customize prompts/timeline_prompts.md or set TIMELINE_PROMPT_FILE to your own file; custom operation prompt path can be set via CUSTOMIZED_PROMPT_FILE. OpenAI compatible: OPENAI_* variables apply to OpenAI official and SiliconFlow, OpenRouter, DeepSeek, etc.; for third-party, change OPENAI_BASE_URL. See User Guide 3.6 Prompts.

Next Steps

• Full configuration and model selection: docs/USERGUIDE.md
• GitHub-only setup details: docs/GITHUB_RELEASE_INBOX.md
• Shortcut templates and manual setup: shortcuts/README.md
• Keep storage when syncing from upstream: docs/STORAGE_MIGRATION.md

License

MIT License