ysyx2008/SailFish

---

Why a Secretary, Not Just a Tool

SailFish isn't a chatbot waiting for commands. It's a personal secretary living on your desktop — with its own rhythm, proactive monitoring, the ability to reach out to you, memory of who you are and what you discussed, and the power to actually operate your computer.

Pain Point	SailFish Approach
Chat-only AI can't act	Your secretary runs commands, reads/writes files, controls the browser, sends email, and makes documents
Every conversation starts from zero	Three-tier memory + knowledge base: knows you, recalls past work, files retrieved when needed
You have to watch and wait	Awaken + Watch: proactive monitoring, scheduled work, push notifications
Tied to your desk	The same secretary via WeChat, DingTalk, Feishu, WeCom, Slack, Telegram, or the web
Don't know the command?	Describe it in natural language; your secretary plans and executes. Risky steps ask first
Intranet / offline	Private models, local Ollama, and proxies

Two Ways to Work, One Workspace

Entry	Relationship	Character
Tasks	You delegate specific work	One-off, parallel, isolated; the desktop mainstay. Queue the next message while a task is still running; a task can carry a goal it keeps working toward on its own
Reach	Your ongoing relationship	Always-on, multi-channel; your secretary can reach out first; work can be handed off as its own task

Tasks come first on desktop. All IM channels flow into one Reach line — not a different bot each time.

The window is three columns: new chat, Reach, Terminal, and recent conversations on the left; the current work in the middle; artifacts on the right, gone when collapsed. Watch, Todos, Contacts, and Memory have their own panels.

Core Capabilities

It remembers

Feature	Description
🧬 Identity & Personality	Your secretary has a character; a birth conversation on first launch, more “yours” the more you use it
🧠 Three-Tier Memory	Knows your habits and facts; recalls past work on demand; view and edit it from the secretary menu
📚 Knowledge Base	Documents stay on your machine, can be organized in folders, and are retrieved when needed

It actually works

Feature	Description
🖥️ Local & SSH Terminal	Operates your machine and remote servers; an SSH session can also work on local files and commands; split panes can mix local and multiple hosts; reconnects after a drop
👥 Sub-agents & Teams	Up to 1,000 unfinished helpers; organize specialists under team leaders, review results, and inspect progress, group conversations, and artifacts in the team overview; resume after interruptions
📁 File Manager	Dual-pane local and remote files, drag-and-drop transfer
🌐 Browser Bridge	Control the Chrome / Edge / Firefox you already have open — logins preserved
📦 Artifacts	WYSIWYG Markdown, live HTML/URL preview, screenshot follow-ups, send to phone; Word / Excel / PDF / WPS open in place
💻 Coding	Edit a local project, check syntax, and roll back by turn; off by default, for advanced users to enable
🛡️ Safety	Command risk levels; dangerous steps ask first; strict / relaxed / free, or “Approve for me”

Secretary work

Feature	Description
👤 Contacts	People and organizations, with an address and how to reach them; people mentioned in conversation are recognized
✅ Todos	Local todos with due-soon reminders and importance sorting; hand one off into a new task
📧 Email	Read, search, write, reply — Gmail, Outlook, QQ Mail, 163, and more
📅 Calendar	Google, iCloud, Outlook, CalDAV — manage the schedule in natural language
📊 Office Skills	Word, Excel, editable PowerPoint, charts, PDF; official-document styles and WPS formats
🐦 Feishu / DingTalk / WeCom	Calendar, tasks, approvals, docs, Bitable, and other workplace skills

It reaches you

Feature	Description
🌅 Awaken	Your secretary keeps its own rhythm and contacts you when it should
👁️ Watch & Sensors	Heartbeat, file-watch, calendar, email, webhooks, and more; a dedicated ops panel so you see problems at a glance
💬 Multi-Channel Reach	WeChat (scan to sign in), DingTalk, Feishu, WeCom, Slack, Telegram, web remote — one secretary
🔊 Speech	Replies can be read aloud; speech recognition is download-on-demand so the installer stays small

It can grow

Feature	Description
🛒 Skill Market & Plugins	One-click community skills; plugins extend tools, providers, and IM channels
🔌 Connectors (MCP)	Databases, APIs, and other external tools; connect when enabled, load on demand
🔍 Web Search	Bocha, Zhipu, Kimi, Tavily, Google, Jina, and more
🖥️ CLI	sailfish runs the full backend headless; share desktop data or isolate with a sandbox
💾 Your data	Full backup and restore; move the data directory to another disk; copy a redacted crash summary
🔒 Always there	System tray and launch-at-login

Quick Start

Download

Get the latest release from the website or GitHub Releases. In mainland China, the site’s Aliyun mirror is usually faster.

Platform	Requirements	Packages
macOS	13+, Intel and Apple Silicon	.dmg (arm64 / x64)
Windows	Windows 10 / 11, Server 2016+	NSIS installer, plus a portable ZIP

No official Linux desktop build. On Windows, if the installer fails (e.g. some Windows Server 2016 hosts), use the ZIP.

First launch on macOS may say the developer cannot be verified: System Settings → Privacy & Security → Open Anyway, or xattr -cr /Applications/SailFish.app.

AI Configuration

SailFish does not ship a model — you bring an API key. Settings include one-click presets; configure several and switch anytime.

{
  "name": "Daily DeepSeek",
  "apiUrl": "https://api.deepseek.com/chat/completions",
  "apiKey": "sk-xxx",
  "model": "deepseek-flash"
}

Recommended models (Function Calling required):

• Daily (fast, affordable, and can see images): DeepSeek V4.1 Flash · Qwen 3.7 Plus · Claude Sonnet 5 · Gemini 3.8 Flash
• Complex (long reasoning, multi-step): DeepSeek V4 Pro · GPT-5.6 · Claude Opus 5 · Grok 4.7
• Vision (screenshots, images, scanned PDFs): DeepSeek V4.1 Flash (enough as the default) · Doubao Seed 2.1 · Qwen 3.7 Plus · GPT-5.6 · Gemini 3.8 Flash · Claude Sonnet 5
• Local / offline: Ollama (preset qwen3.5:9b; small context windows struggle with multi-step work)
• More: Zhipu GLM-5.3 Flash, Kimi K3, MiniMax M3, and any OpenAI-compatible endpoint

A text-only model can be paired with a vision model: pictures go there, then conversation returns. For web search in China, Bocha, Zhipu, or Kimi all work. Step-by-step: First setup.

Development

npm install
npm run dev              # development
npm run build:mac        # macOS package
npm run build:win        # Windows package
npm run test             # unit tests
bash electron/cli/test-cli.sh --no-ai   # CLI regression (no model calls)

Dev Windows build (for the author’s cross-platform testing only): npm run build:win:remote triggers a GitHub Actions build; the artifact overwrites a fixed OSS path in 8–10 minutes. Stability is not guaranteed. https://sfterm-download.oss-cn-wuhan-lr.aliyuncs.com/dev/SailFish-Setup-dev.exe

See CONTRIBUTING.md. Day-to-day work is on develop; open PRs against develop, not main.

CLI

After install, add the sailfish command from Settings → Data Management. In a checkout, use npm run sailfish or npm run install:cli.

sailfish "list markdown files in this folder"
sailfish --sandbox "try this without touching desktop data"
sailfish models
sailfish --help

By default it shares desktop data and asks before dangerous steps (--mode relaxed). --sandbox isolates runtime data while still borrowing AI profiles and credentials. --mode free / --free skips confirmations — use with care.

Running from the repo defaults to the sandbox so desktop history is not overwritten.

Architecture

The desktop app (Electron) and the CLI share one backend. The UI renders; sessions, memory, and tool execution live in the backend. IM and web remote use that same backend — not a separate bot.

Documentation

For users

• What is SailFish
• Install
• User guide
• Skill market
• Data safety
• Changelog · Website changelog

For developers and integrators

• How the Agent works
• IM integration
• Plugin development
• Contributing

License

Dual licensing: AGPL v3.0 for open source use; commercial license available.

• ✅ Personal use, research, education
• ✅ Medical and non-profit organizations
• ✅ Enterprise internal use (≤1000 installations; modifications must be open-sourced under AGPL)
• 💼 Commercial license required for: >1000 installations, SaaS / product embedding, closed-source modifications, changing the logo or name, or removing “Support the author”

See LICENSE.

Links

• 🌐 Website
• 📦 GitHub
• 🐛 Issues
• 💬 QQ group: 1078041072

Acknowledgements

Built with Electron, Vue.js, xterm.js, LanceDB, and many other open source projects.