🤖 AideAgent

The open-source Claude Desktop alternative — 28 tools, local RAG, Prompt Caching, all in one cross-platform desktop app.

👀 What is AideAgent?

AideAgent is a local-first desktop AI assistant that does everything Claude Desktop does — and more — but free and open-source (Apache-2.0).

🛡️ Your data never leaves your machine — offline MiniLM-L6 embedding, OS-encrypted API keys, SSRF protection
⚡ Save 90% on API bills — explicit Prompt Caching (Anthropic + DeepSeek) with live hit-rate display
🧰 28 built-in tools — file/shell/git/GitHub/MCP/LSP/skill/kb, all in one binary
🔌 Plays nice with the ecosystem — auto-detects Claude Code/Desktop MCP configs, one-click install of 200+ skills
💬 WeChat bot — talk to your agent from your phone

Built by one developer in 24 days. 17 releases. 180 commits. 25 test files. ~200MB desktop app.

📸 Preview

Main Chat Interface
左侧会话列表 + 工作目录，中间流式对话区，底部 4 个智能开关（计划模式 / 知识库 / 联网搜索 / 深度推理）

Knowledge Base Search (Hybrid RAG)
FTS5 关键词 + 向量语义 + RRF 融合，3 亿字 Obsidian vault 也能秒级定位

28 Tools Overview
Agent 自主介绍能力：编程、搜索、管理文件、操作 Git、查阅知识库，一应俱全

Local-First Tech Stack
内嵌 MiniLM-L6 离线嵌入 / 198 篇已索引笔记 / 4 类持久记忆 / 0 远程依赖

⚖️ AideAgent vs The Alternatives

Tired of paying $20/mo for tools that don't have RAG, hooks, or WeChat integration?

Feature	AideAgent	Claude Desktop	Cursor	Cline	Continue
Price	Free + OSS	$20/mo	$20/mo	Free + OSS	Free + OSS
Open Source	✅ Apache-2.0	❌	❌	✅ Apache-2.0	✅ Apache-2.0
Local RAG / Knowledge Base	✅ Built-in hybrid	❌	⚠️ Paid add-on	❌	⚠️ Basic
Prompt Caching	✅ Explicit + visible	✅ Built-in (hidden)	⚠️ Black box	❌	❌
Hook System	✅ 3 events	❌	❌	❌	❌
MCP Support	✅ Auto-detect	✅	⚠️ Limited	✅	⚠️ Limited
LSP Integration	✅ Built-in	❌	✅	❌	⚠️ Basic
WeChat Bot	✅	❌	❌	❌	❌
Self-Evolving Memory	✅ Auto	❌	❌	❌	❌
Auto-Compress & Continue	✅ Up to 5×	⚠️ Limited	⚠️ Limited	⚠️ Limited	❌
Skill Ecosystem	✅ 200+	⚠️ Limited	⚠️ Limited	❌	⚠️ Limited
Offline Embedding Model	✅ Bundled	❌	❌	❌	❌
API Key Encryption	✅ OS-level	✅	✅	✅	⚠️ Local file
Install Size	~200MB	~200MB	~400MB	~80MB	~60MB
Platforms	Win/Mac/Linux	Win/Mac	Win/Mac	VSCode	VSCode/JetBrains

Verdict: If you want Claude Desktop's polish plus RAG plus Hooks plus WeChat — without the $20/mo tax — AideAgent is for you.

🚀 Quick Start (30 seconds)

Option 1: Download Pre-built Binary (Recommended)

👉 Download for Windows / macOS / Linux →

No Node.js required. Double-click and go.

Option 2: Run from Source

git clone https://github.com/quanzefeng/AideAgent.git
cd AideAgent/desktop
npm install        # postinstall auto-downloads MiniLM-L6 (~23MB)
npm start

3-Step Setup

Open Settings → API Config → choose a provider (DeepSeek / Claude / GLM / Qwen / Ollama / ...)
Paste your API key → saved with OS-level encryption (Windows DPAPI / macOS Keychain / Linux libsecret)
Start chatting → Agent has 28 tools ready to go

💡 No API key? Use Ollama / LM Studio / llama.cpp for fully local inference. Zero data leaves your machine.

🌟 Core Features

🛡️ Local-First Privacy

Offline MiniLM-L6 embedding model bundled (384-dim, ~23MB)
OS-level encryption for all API keys (Electron safeStorage)
SSRF protection — web_fetch blocks 127.0.0.1 / 192.168.* / 169.254.*
Path traversal prevention in KB writes and Hook scripts
Command injection proof — git/gh use array-based spawn, never shell strings
XSS proof — KB panel uses textContent, zero innerHTML
Dangerous command detection — rm -rf /, format, etc. require explicit confirmation

⚡ Prompt Caching (Save 90% on API bills)

Stable system prompt — dynamic content (memory, KB, tasks) moved to user messages, system prompt never changes → always cacheable
Lazy KB loading — KB results injected only on first turn
Anthropic explicit cache — cache_control markers on system + tools + history, 1-hour TTL
DeepSeek auto-cache — message ordering optimized for maximum prefix match
Live hit-rate display — 💾 90% cached shown after each reply with color coding (🟢≥80% / 🟡≥50% / 🔴<50%)

🧰 28 Built-in Tools

Category	Tools
Files & Code	`file_read`, `file_write`, `file_edit`, `grep`, `glob`, `lsp` (go-to-def, references, hover)
Shell	`bash` (cross-platform pwsh/bash, dangerous command detection, Hook interception)
Web	`web_search` (Tavily + meta-search fallback), `web_fetch` (SSRF-protected)
Version Control	`git_diff`, `git_commit`, `git_branch`, `gh_pr`, `gh_issue`, `gh_repo`
Knowledge	`kb_search` (FTS5 + vector + RRF), `kb_write`, `kb_get_note`, `write_memory`
Tasks	`TaskCreate`, `TaskUpdate`, `TaskList`, `TodoWrite`, `AskUserQuestion`
Skills	`skill`, `invoke_skill`, `create_skill` (L2/L3 dual system)
Sub-Agent	`Agent` (read-only parallel research)

🔌 MCP + Skill Ecosystem

MCP — stdio + HTTP transports, auto-detects Claude Code/Desktop/OpenCode configs
200+ skills installable via npx skills add (L3 system)
L2 agent-created skills — agent learns your workflow and creates reusable skills automatically
Curator auto-archives skills unused for 30+ days, detects duplicates

🧠 Hybrid RAG (Obsidian-Native)

FTS5 keyword search + vector semantic search (MiniLM-L6 or Ollama)
RRF fusion — merges both result sets for optimal relevance
MRL compression — auto-detects Matryoshka-capable models (e.g. qwen3-embedding), compresses 1024→384 dim losslessly
Smart truncation — auto-detects Ollama model context length
Instant injection — relevant notes auto-injected into system prompt

🪝 Hook System (Programmable Agent Behavior)

3 events — PreToolUse / PostToolUse / SessionEnd
JSON protocol over stdin/stdout, cross-platform
Safety guards — enabled toggle, tools filter, path traversal protection, timeout fallback
Built-in examples — dangerous command interception, auto-formatting, session notifications

💾 Self-Evolving Memory

Post-session reflection — agent analyzes recent exchanges, extracts PREFERENCE / DECISION / KNOWLEDGE
Auto-save to USER.md / MEMORY.md — zero user friction
Semantic selection — auto-picks most relevant memories per conversation
Deduplication — prevents saving redundant content

🔄 Auto-Compress & Continue

Context-aware trigger at 90% token usage (256K window)
LLM summarization — preserves key decisions and file references
Seamless resume — "Auto-continued (N/5)" shown inline
Up to 5× continuations = 50 tool calls × 5 = 250 effective rounds

📁 Project Context (AGENTS.md)

Multi-standard — reads AGENTS.md first, CLAUDE.md as fallback
Two-level loading — project root + ~/.aideagent/CLAUDE.md
Zero config — auto-loaded if present

💬 WeChat Bot

QR-code login to WeChat
Send messages to your agent from your phone
Full conversational capabilities, end-to-end

🎨 UI / UX

Dark theme, streaming responses, expandable reasoning
LaTeX math + code syntax highlighting (highlight.js)
One-click Chinese/English UI switch
12 settings panels (API / Avatar / Skills / Prompt / MCP / Social / Memory / KB / Skills / Font / Language / About)
Custom avatar and agent name

🏗️ Architecture

desktop/
├── main.mjs                 # App entry
├── preload.cjs              # IPC bridge (~60 channels)
├── core/
│   ├── agent-loop.mjs       # Conversation loop + auto-continue + auto-review
│   ├── tool-executor.mjs    # 28 tools + cross-platform shell
│   ├── tool-definitions.mjs # Tool schemas
│   ├── system-prompt.mjs    # Prompt builder + AGENTS.md loading
│   ├── format-adapters.mjs  # OpenAI/Anthropic adapters + Prompt Caching
│   ├── token-budget.mjs     # Token estimation + context compression
│   ├── hook-manager.mjs     # Hook system
│   ├── sub-agent.mjs        # Read-only sub-agents
│   ├── state.mjs            # Shared state + platform detection
│   ├── ipc-handlers.mjs     # IPC handlers (~80)
│   ├── memory-selection.mjs # Semantic memory selection
│   ├── skill-scanner.mjs    # L3 skill scanner
│   └── workspace-config.mjs # Workspace persistence
├── session-db.mjs           # Session DB (SQLite + FTS5)
├── memory-store.mjs         # Memory store
├── skills-store.mjs         # L2 skill store
├── knowledge-store.mjs      # KB RAG (SQLite + FTS5 + vectors)
├── mcp-manager.mjs          # MCP protocol (JSON-RPC 2.0)
├── lsp-manager.mjs          # LSP client
├── update-manager.mjs       # Auto-updater
├── renderer/                # UI (vanilla JS, no framework bloat)
│   ├── index.html
│   ├── app.js
│   ├── style.css
│   ├── translations.js      # i18n (zh-CN + en)
│   └── modules/
├── test/                    # 25 Vitest test files
├── models/                  # Bundled MiniLM-L6
└── scripts/                 # Build scripts

Stack: Electron 40 · Node.js 22 · SQLite (FTS5 + vectors) · Vitest · Playwright · electron-builder

No TypeScript build step — vanilla JS keeps startup fast and bundle small.

🔬 Who's Using AideAgent?

Be the first to star ⭐ the repo and open an issue to get featured here.

📚 Documentation

📖 User Guide — Detailed walkthrough of every feature
🛠️ Developer Guide — Architecture deep-dive, contributing
🔌 MCP Setup — Connect external tool servers
🪝 Hook Examples — Pre/Post/SessionEnd recipes
🎯 Prompt Caching — How the cache works + cost benchmarks
🤖 Multi-Model Setup — DeepSeek, Claude, GLM, Qwen, Ollama, ...

🗺️ Roadmap

Agent Marketplace — share and install community-created skills
VSCode Extension — same agent, in your editor
Multi-Agent Collaboration — orchestrate multiple agents on one task
Voice Mode — push-to-talk with Whisper integration
Cloud Sync (optional, E2E-encrypted) — for users who want it

See open issues for the full list.

🤝 Contributing

Contributions are welcome! See CONTRIBUTING.md for the full guide.

Quick start:

git clone https://github.com/quanzefeng/AideAgent.git
cd AideAgent/desktop
npm install
npm run dev      # development mode with hot reload
npm test         # run Vitest suite
npm run lint

Good first issues: good-first-issue label

📜 License

Apache License 2.0 — free for commercial and personal use.

💖 Sponsorship

If AideAgent saved you time or money, consider buying the author a coffee ☕

📬 Contact

💬 WeChat: q2993919594
📧 Email: zefengquan5@gmail.com
🐙 GitHub: @quanzefeng

If this project helps you, please consider starring it — it helps more developers find it.

Made with ❤️ by one developer in 24 days.

Name		Name	Last commit message	Last commit date
Latest commit History 184 Commits
.aideagent/hooks		.aideagent/hooks
.github/workflows		.github/workflows
.planning/codebase		.planning/codebase
assets		assets
desktop		desktop
docs/screenshots		docs/screenshots
kb-test-backup		kb-test-backup
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md

Folders and files

Latest commit

History

Repository files navigation

🤖 AideAgent

The open-source Claude Desktop alternative — 28 tools, local RAG, Prompt Caching, all in one cross-platform desktop app.

👀 What is AideAgent?

📸 Preview

⚖️ AideAgent vs The Alternatives

🚀 Quick Start (30 seconds)

Option 1: Download Pre-built Binary (Recommended)

Option 2: Run from Source

3-Step Setup

🌟 Core Features

🛡️ Local-First Privacy

⚡ Prompt Caching (Save 90% on API bills)

🧰 28 Built-in Tools

🔌 MCP + Skill Ecosystem

🧠 Hybrid RAG (Obsidian-Native)

🪝 Hook System (Programmable Agent Behavior)

💾 Self-Evolving Memory

🔄 Auto-Compress & Continue

📁 Project Context (AGENTS.md)

💬 WeChat Bot

🎨 UI / UX

🏗️ Architecture

🔬 Who's Using AideAgent?

📚 Documentation

🗺️ Roadmap

🤝 Contributing

📜 License

💖 Sponsorship

📬 Contact

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases 17

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages