Language: 繁體中文

在你電腦上的個人 AI 代理

Instrumentarium

The agent can use the terminal, files, web, browser, memory, and more. Most power lives in skills and plugins so the core stays lean. The matrix below is a novice bridge: job → start surface → toolsets.

第一次來？從場景或能力目錄開始。 使用場景 · 能力

Back to capabilities 生態 Getting started
Command Center
按 ⌘. / Ctrl+.，從會話搜索跳到存儲健康與令牌用量，無需離開 Desktop。

產品長什麼樣

帶 Plan · Report · Design · Runs 的 Desktop 聊天、Work 收件箱、審批與 skills。

畫面截至 2026-08-06 — 隨每個 Desktop 版本更新。

帶 Plan 欄的 VibeOS Desktop 聊天
聊天 + Plan 欄 — 提問、計劃、行動、驗證。
Work Mode 收件箱
Work 收件箱在聊天窗口關閉後仍讓目標可見。
審批門
敏感操作會等待明確批准。
Skills 選擇器
Skills 加載可復用劇本，而不膨脹核心。

畫面爲注重隱私的產品截取（非客戶實時會話）。

從哪裡開始

新手橋：將作業與起始表面相匹配，然後啓用您需要的工具集。在要求新的核心工具之前優先選擇技能/插件/MCP。

工作 從這裡開始 工具組和邊緣
船舶代碼Desktop聊天+Plan rail，或CLI/TUI終端·文件·代碼執行·技能（lang/PR）
研究簡介Desktop + Research Pro 預設網絡·瀏覽器·研究技能·可選MCP
運營文摘Desktop 或 Work Mode 目標終端 · 網頁 · cron · 批准
消息代理自託管網關消息傳遞工具集 · 批准 · 平臺適配器
常備自動化Cron 作業或 kanban 板cronjob · kanban · 委派（進程本地）
定製能力生態系統頁面 → 足跡階梯技能 → 插件 → MCP → 核心工具（最後）

使用「vibeos tools」或 config.yaml 啓用/禁用每個平臺。服務門控工具（kanban、Home Assistant 等）在配置之前不會出現在架構中。

What it can touch

Terminal & environments

Local shell, Docker, SSH, and cloud backends. Background processes with completion notify.

Files & code

Read, search, patch, and execute code; project context from AGENTS.md / CLAUDE.md / SOUL.md.

Web & browser

Web search/extract (Firecrawl, Exa, Tavily, Yandex Search, and more), browser automation, `vibeos research` workspaces/presets, and cited briefing skills.

Memory & session search

Write curated memory, recall prior sessions, optional external memory providers.

More tool families

Built-in tool families

Cron, kanban, delegate

Schedule jobs, manage multi-agent boards, spawn leaf or orchestrator subagents.

Vision, image, TTS

Multimodal analysis, image generation, speech synthesis — enabled when providers are set.

Research Pro

Cited multi-source research without growing the core tool schema: search backends, skills that grade evidence, MCP for apps, and scheduled briefings.

Search & extract

Yandex Search (RU), Exa, Tavily, Firecrawl, and more via `web_search` / `web_extract`. Preset: `vibeos research preset research-pro --apply` (or Desktop → Settings → Research Pro).

Research skills

Bundled: professional-research, source-critique, comparative-analysis, academic-literature, longform-synthesis, quantitative-analysis, analysis-orchestration, visual-source-extract. Optional RU registries: `vibeos skills install official/research/ru-public-sources`.

Workspaces & CLI

`vibeos research run|list|show` scaffolds dated workspaces under ~/.vibeos/research/ (prompt + sources template — no LLM call). Modes: brief, deep, academic, compare, quant…

Apps & literature MCP

Catalog: Notion, GitHub, Obsidian, Linear, Tavily map/crawl, Google Workspace / Gmail / Drive, Perplexity, scholar-search. Install with `vibeos mcp install <name>`. Cron briefings: https://vibeos.com.ru/handbook/guides/research-daily-briefing/.

How capability grows

Skills

Procedural docs + scripts the model loads on demand. Zero permanent tool-schema cost.

MCP servers

Catalog MCP servers (Linear, Notion, GitHub, Google Workspace, Perplexity, scholar-search, Obsidian, n8n, …) — install with `vibeos mcp` without shipping SaaS SDKs in core.

Plugins

Register tools, hooks, CLI commands, model providers, and memory backends from ~/.vibeos/plugins.

Model providers

OpenRouter, Anthropic, OpenAI-compatible endpoints, and more — BYOK on hosted plans.

Hooks

Run custom code at lifecycle points — observe, rewrite, or block — without crashing the agent if a hook fails.

Control surface

Enable or disable toolsets per platform with `vibeos tools` or config.yaml. Security approvals, checkpoints/rollback, and credential pools stay under your control.