EvoLoop Agent

Universal Agent System · One Person, One Digital Fleet

Releases  ·  Website  ·  Try Online  ·  Issues

English | 中文 | 日本語 | 한국어

---

EvoLoop is a cross-platform, long-horizon, collaborative universal agent system. It plans complex tasks autonomously, controls browsers, desktops, and external devices directly, and remembers historical experience — getting better the more you use it.

From personal computers to Linux servers, from single devices to distributed multi-agent networks, EvoLoop evolves AI from a "chat box" into an autonomous employee that can execute and close loops in real business workflows.

Desktop

Mobile

---

🌟 Highlights

Capability	Description
Long-Horizon Execution	300+ continuous steps per task, stable for hours, with automatic error recovery
A2A Multi-Agent Collaboration	Agents across devices discover, delegate, and return results asynchronously
Cross-Device Swarm Control	Phones and computers control each other; one PC can manage multiple Android devices
Imitation Learning	Generate reusable automation skills from screen recordings or operation traces
Chat + Terminal Dual Mode	Natural language chat and native terminal in one unified desktop interface
Server / Edge Deployment	Deploy as a persistent node on Linux servers, cloud hosts, or embedded devices
Graph-RAG Episodic Memory	Archive experiences automatically and avoid repeating the same mistakes
Dynamic Tools + MCP	Create Python tools at runtime and mount MCP services natively
Deep Code Understanding	Tree-sitter + vectors + code graph for navigating large codebases
Human-in-the-Loop	High-risk actions require human confirmation; core data stays local by default

---

🏗 System Architecture

graph TB
    subgraph "User Layer"
        WEB["🌐 Web App\nReact + Vite"]
        DESKTOP["💻 Desktop App\nTauri + React"]
    end

    subgraph "Backend Layer"
        API["⚡ FastAPI Server\nREST + SSE"]
        ENGINE["🧠 Multi-Agent Engine\nSupervisor + Worker"]
        WORKER["📬 Async Worker\nHuey / Celery"]
    end

    subgraph "Tool Layer"
        BROWSER["🌐 Browser\nPlaywright"]
        DESKTOP_TOOLS["🖥️ Desktop\nNative System API"]
        MOBILE["📱 Mobile\nADB Driver"]
        MCP["🔌 MCP Client"]
    end

    subgraph "Data Layer"
        PG[("PostgreSQL / SQLite")]
        VECTOR[("LanceDB / pgvector")]
        NEO4J[("Neo4j / File Graph")]
        SEARCH[("Meilisearch / FTS5")]
        REDIS[("Redis / In-Memory Queue")]
    end

    subgraph "Cloud"
        EVOCLOUD["☁️ EvoCloud"]
    end

    WEB --> API
    DESKTOP --> API
    EVOCLOUD -.->|WebSocket| API

    API --> ENGINE
    ENGINE --> WORKER
    ENGINE --> BROWSER
    ENGINE --> DESKTOP_TOOLS
    ENGINE --> MOBILE
    ENGINE --> MCP

    ENGINE --> PG
    ENGINE --> VECTOR
    ENGINE --> NEO4J
    ENGINE --> SEARCH
    ENGINE --> REDIS
    WORKER --> REDIS

Loading

The mobile app is part of EvoLoop's commercial offering and is not included in this open-source repository. This repo controls external mobile devices via ADB.

---

🚀 Quick Start

1. Clone the repository

git clone https://github.com/wonderful-team/evoloop.git
cd evoloop

2. Choose an environment template

# Desktop / local development (embedded, zero middleware)
cp .env.prod.desktop .env

# Web single-user server (embedded)
# cp .env.prod.web.single .env

# Web multi-user production (requires PostgreSQL/Redis/Meilisearch/Neo4j)
# cp .env.prod.web.multi .env

3. Start the development environment

./deploy/dev.sh

Once started, visit http://localhost:20160/docs for the API docs.

4. Server deployment

# One-line Docker install for Linux / macOS
./deploy/install.sh

# CentOS / RHEL / Baota panel
./deploy/install_centos.sh

# Remote Linux server deployment
./deploy/deploy.sh --host=your-server-ip --env-file=.env.prod.web.multi

For more details, see deploy/README.md.

---

🖥️ Deployment Modes

Mode	Scenario	Stack	Config File
Desktop Embedded	Personal computer, local use	SQLite + LanceDB + Huey	.env.prod.desktop
Web Single-User	Personal server	SQLite + LanceDB + Huey	.env.prod.web.single
Web Multi-User	Team / enterprise production	PostgreSQL + Redis + Meilisearch + Neo4j + Celery	.env.prod.web.multi

---

🤖 Models

EvoLoop connects to mainstream LLMs through EvoCloud, and you can switch models with one click in the Web console — no manual config editing required.

Type	Representative Models
Chat	GPT-4o, Claude, Kimi, DeepSeek, Qwen, Doubao
Vision	GPT-4o Vision, Claude 3.5 Sonnet, Kimi Vision
Embedding	nomic-embed, BGE, OpenAI Embedding
ASR/TTS	FunASR, paraformer-zh, System TTS

---

🧠 Memory & Knowledge

• Three-tier memory: short-term conversation context → long-term experience archive → project hot memory (MEMORY.md)
• Automatic distillation: extracts valuable information after each task to avoid repeated mistakes
• Project profile: auto-detects project structure, generates tech-stack summaries, and builds code graphs
• Knowledge capture: captures concepts, decisions, and SOPs from conversations and code changes

---

🧩 Tools & Skills

Tools: file I/O, terminal, browser, web search, code retrieval, scheduler, memory recall, MCP, and more.

Skills: generate reusable skills from natural language descriptions or screen recordings, with parameterization and one-click invocation.

---

📦 Build & Release

# Desktop
./deploy/build.sh macos-arm64 --env-file=.env.prod.desktop
./deploy/build.sh macos-x86_64 --env-file=.env.prod.desktop
./deploy/build.sh windows --env-file=.env.prod.desktop

# Web static assets
./deploy/build.sh web --env-file=.env.prod.web.multi

# Model download
./deploy/download_models.sh --models nomic-embed,paraformer-zh

---

📋 Changelog

v0.7.0 — Cross-device swarm control, long-horizon execution, Chat + Terminal dual mode, imitation learning, server/edge deployment, A2A multi-agent collaboration.

For full history, see Releases.

---

🤝 Community & Support

• Issues: GitHub Issues
• Docs: Wiki
• Email: wonderful@develop-assistant.cn

---

⚠️ Disclaimer

1. This project is licensed under the MIT License for technical research and learning.
2. Agent mode consumes significantly more tokens than regular chat — please monitor costs. The Agent can access your local operating system, so use it only in trusted environments.
3. High-risk operations trigger human confirmation. Please keep your EvoCloud credentials and local data secure.