Apache-2.0 開源、本地優先的 AI 桌面助手，內建 29 個技能、MCP、排程與觸發器；DeepSeek Harness 整合在路線圖中，穩定版暫用 Abu 原生 harness。
DSH 整合
生態系相關
作者聲明
安全稽核
未稽核
最後核實
2026-08-21
授權條款
Apache-2.0
01它能幫你完成什麼？
Delegate everyday office work — organizing files, drafting docs, building reports — to a locally-run AI desktop agent
Files read/written, commands executed, documents and reports produced on your machine by an autonomous planning agent
Individuals and office workers who want a privacy-first local AI agent instead of cloud-only chat
Automate recurring workflows with cron schedules and event triggers (file changes, HTTP webhooks, IM messages)
Scheduled tasks (e.g., daily AI news digest) and trigger-driven agent runs, with missed runs replayed on next launch
Teams and power users who want hands-off workflow automation on their desktop or as an IM bot (Lark/DingTalk/WeCom/Slack)
Produce office deliverables — PDF, PPTX, DOCX, Excel, infographics — from chat prompts
Generated documents and designs via 29 built-in skills, plus self-evolving custom skills crystallized from your own workflows
Knowledge workers who need finished documents (weekly report PPT, Excel from a PDF) rather than raw text output
02如何將外掛接入 DeepSeek Harness？
先決條件
- macOS (Apple Silicon / Intel) or Windows x64
- An API key for a cloud AI provider (e.g., DeepSeek, Anthropic), or Ollama installed for local models
安裝步驟
- 01
Download the installer for your platform from GitHub Releases (`Abu-x.x.x-mac-arm64.dmg`, `Abu-x.x.x-mac-x64.dmg`, or `Abu-x.x.x-windows-x64-setup.exe`) and run it
- 02
Open Abu → Settings → AI Service Management: choose a provider (e.g., DeepSeek, Anthropic), enter your API Key, click verify — or use zero-config Ollama local models / any OpenAI-compatible custom endpoint
- 03
Return to the main screen, use the model selector to pick your preferred model, and start chatting
驗證整合成功
作者未說明
03DSH 整合程度與能力邊界
Standalone local-first AI desktop assistant; DeepSeek Harness is among the first targets of the in-progress multi-harness runtime integration — stable releases currently run Abu's native harness
Autonomous agent with Plan Mode
natural-language tasks in chat→end-to-end task completion: planning, tool invocation, file read/write, command execution
reads/writes local files and executes shell commandshigh-risk steps (delete / overwrite / send / install) are gated behind Plan Mode confirmation; per-conversation permission modes (Request Approval / Smart Review / Full Autonomy)29 built-in skills + MCP connectors
skill prompts (PDF/PPTX/DOCX/Excel, frontend design, diagrams, browser automation) and MCP servers→skill-driven deliverables; external services (databases, search engines, GitHub, Playwright) connected via Model Context Protocol
skills are installed from the in-app Skill StoreMCP connectors reach external services (GitHub, Playwright, databases) over the networkSchedules, triggers & IM bot
cron plans; file-watcher globs; HTTP webhooks; IM messages (Lark/DingTalk/WeCom/Slack)→agents invoked automatically on schedule or events; @Abu team bot in chat channels
auto-generated POST endpoints for HTTP webhooksagents run automatically without manual prompting, under a four-level trigger permission modelLocal file-based memory with OS sandbox
conversations and learned preferences→three-tier memory (personal `~/.abu/memory/`, per-project memory, ABU.md rules) persisted as markdown files
writes memory/rule files under `~/.abu/` and `{workspace}/.abu/`shell commands run inside macOS Seatbelt / Windows ConstrainedLanguage sandbox with network isolation (local proxy + domain whitelist)
04適合誰？何時不該用？
適合
- Individuals and office workers who want a privacy-first local AI agent instead of cloud-only chat
- Teams and power users who want hands-off workflow automation on their desktop or as an IM bot (Lark/DingTalk/WeCom/Slack)
- Knowledge workers who need finished documents (weekly report PPT, Excel from a PDF) rather than raw text output
不適合
- DeepSeek Harness support is roadmap-only at this time: multi-harness integration is in progress with DeepSeek Harness among the first targets, but stable releases currently run Abu's native harness — there is no dsh install path yet.
05相容性、維護與安全提醒
- DeepSeek Harness support is roadmap-only at this time: multi-harness integration is in progress with DeepSeek Harness among the first targets, but stable releases currently run Abu's native harness — there is no dsh install path yet.
- The Windows installer is current-user only and Authenticode-unsigned, so SmartScreen may require "More info → Run anyway"; only macOS builds are signed and notarized.
- Packaged apps ship for macOS (Apple Silicon / Intel) and Windows only; other platforms require building from source (Node.js 24 + Rust + platform build tools).
Apache-2.0 · actively maintained (latest release v0.40.0, 2026-08-20)
06常見問題
Abu 如何接入 DeepSeek Harness？
目前尚未正式支援。README 說明多 Harness 整合正在開發中，DeepSeek Harness 是首批整合目標之一；目前穩定版使用 Abu 自研的原生 harness，因此現階段 Abu 以獨立桌面應用程式形式執行，而非在 dsh 內執行。
支援哪些平台？怎麼安裝？
支援 macOS（Apple Silicon / Intel）和 Windows x64。從 GitHub Releases 下載對應的 dmg/setup.exe 安裝即可；macOS 版本已簽署並公證，Windows 安裝包未做 Authenticode 簽署，SmartScreen 提示時需點選「更多資訊 → 仍要執行」。
可以用哪些 AI 模型？
支援 12+ 雲端供應商（DeepSeek、Anthropic Claude、OpenAI、通義千問/百煉、豆包、月之暗面、智譜等），可零設定接入 Ollama 本機模型，也相容任何 OpenAI 相容或 Anthropic 相容的自訂端點。
我的資料會上傳嗎？
Abu 是本地優先：資料與 API Key 都儲存在本機（Key 加密儲存），不經過第三方伺服器。Shell 指令在作業系統沙箱內執行，並透過本機代理加網域白名單實現網路隔離。
是免費開源的嗎？
是，Apache License 2.0，可免費使用、修改與散布（含商業用途）。企業版（團隊協作、SSO、稽核日誌、私有化部署）需要另行授權。
07相關的 DSH 工作流程
ruflo
作者 ruvnet
原版智慧體元框架（meta-harness）：支援多智慧體叢集部署、自主工作流編排與對話式 AI 系統構建，內建自適應記憶、自學習智慧與 RAG 檢索，原生整合 Claude Code / Codex / Hermes 等大量工具。
reactive-resume
作者 amruthpillai
注重隱私的開源簡歷生成器，安全、可定製、可移植，永久免費，支援自託管。
everos
作者 evermind-ai
為所有 AI 智慧體提供的便攜統一記憶層，本地優先、Markdown 原生、資料歸使用者所有，跨應用與工作流自進化。
yao
作者 yaoapp
在桌面、移動端、瀏覽器或 API 上統一管理你的智慧體與工作區，看板式任務追蹤，支援自託管。
08資料與來源
with [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) among the first integration targets. Stable rel…
此頁面根據專案公開文件、儲存庫中繼資料與 DSH Plugins 的結構化解析所產生；最後核實於 2026-08-21。發現錯誤？提交更正。
最佳 DeepSeek Harness 外掛
從全目錄挑出的 12 個值得優先安裝的外掛，涵蓋各個分類。