flock Documentation

Flock

A desktop multi-agent harness built with Rust, Tauri, and React, powered by langgraph-rust.

Visual Workflow | Multi-Agent Harness | Built-in Agent | Safe Sandbox & VNC | Cross-Platform | Any API Key

English | 简体中文

Note: This project is built on top of langgraph-rust, which is my personal Rust implementation of the LangGraph framework.

Refactoring History: Flock has been completely rewritten from the ground up. The original version was a Python-based application using LangGraph, LangChain, and FastAPI as the backend. The current version is a native desktop application with a Rust backend, powered by Tauri for the desktop shell. This rewrite brings significant improvements in performance, reliability, and user experience.

Legacy Code: The original Python codebase is preserved in the legacy/python branch for reference.

📋 Quick Navigation Copy Chapter

Multi-Agent Harness & Workflow · Core Features · Architecture · Quick Start

🤝 Multi-Agent Harness & Workflow — Visual Orchestration for AI Agents Copy Chapter

Flock is more than just a chat client. It's a comprehensive multi-agent harness combined with a powerful Visual Workflow Editor. It orchestrates and runs AI agents locally or inside sandbox containers to read/write files, execute bash commands, browse the web, and run complex pipelines. You see everything the agent does, and you're always in control.

🌟 Core Features Copy Chapter

🕸️ Visual Workflow Editor (ReactFlow Engine) Copy Section

• Drag-and-Drop Canvas: Design complex pipelines visually by connecting custom agent and logic nodes.
• 10 Node Types Shipped:
  • start & answer: Define workflow inputs and final delivery parameters.
  • llm & agent: Pure inference nodes and tool-enabled LangGraph agents.
  • classifier & ifelse: Semantic routing routers and conditional branch checkers.
  • code: Custom JS/Python code runners to execute arbitrary transformations.
  • human: Interactive interrupts asking for manual feedback or input.
  • plugin & parameter_extractor: Expose custom tools/plugins and extract structured data.
• Version Control & Execution History: Easily manage multiple versions of your visual workflows, track historical execution paths, and debug states incrementally.

👥 Human-in-the-Loop (HITL) & Safe Interaction Copy Section

• Interactive Tool Approvals: Crucial actions (like writing files, running bash scripts, or modifying configurations) require explicit user approval before execution. You can inspect the plan, approve, or deny on the fly.
• VNC Screen Takeover: Directly interact with sandboxed browser and OS interfaces. When the agent gets stuck on captchas, you can take over mouse/keyboard controls seamlessly.
• Workflow Breakpoints (human node): Place manual intervention checkpoints in the visual flow to request human decisions, form submissions, or variable modifications before proceeding.

🌐 Browser Automation & Computer Use Copy Section

• Playwright Browser: Scrape web data, click links, and bypass forms using an isolated browser execution engine.
• Computer Use: Command virtual desktops using OS-level commands (xdotool). The agent monitors framebuffer screen feeds and handles mouse/keyboard events.

🤖 Built-in Agent — Zero Configuration Copy Section

• Out of the box: No external CLI tools to download or configure. Paste any API Key (OpenAI, Gemini, Anthropic Claude, AWS Bedrock, or Ollama/local) and start immediately.
• Skill Extensions: Extensible via YAML-frontmatter prompts that support Hot-Reloading on file changes.

🔒 Safe Sandbox Environment Copy Section

• Isolated Environment: Executes risky shell scripts and code execution safely inside isolated environment containers.
• VNC Streaming: Stream sandbox desktops directly into the UI via WebSockets for full visual inspection of the running environment.

🔌 MCP & Scheduled Tasks Copy Section

• MCP (Model Context Protocol): Connect to third-party MCP servers once; all assistants and workflows automatically inherit the newly exposed schemas and tools.
• Scheduled Cron Tasks: Set it once, the agent executes scheduled tasks on autopilot. Supports standard cron syntax for automated system maintenance, daily file aggregation, or report generation.

🏗️ Architecture Copy Chapter

graph TD
    UI[Tauri Desktop UI - React + Mantine] <-->|IPC / Tauri Command| TauriHost[src-tauri Core]
    TauriHost <-->|IPC Interface / JSON Protocol| AgentEngine[flock-agent - LangGraph Core]
    AgentEngine -->|Graph State Machine| StateGraph[langgraph-rust]
    AgentEngine <-->|SQLite Checkpointer| DB[(SQLite Database)]
    AgentEngine <-->|Discovery / Frontmatter| Skills[flock-skills]
    
    AgentEngine <-->|Tool Execution| Registry[flock-tools - Tool Registry]
    Registry <-->|Local Host| Builtin[Built-in Tools: Bash, Files, Grep]
    Registry <-->|Model Context Protocol| MCP[MCP Server Registry]
    Registry <-->|Isolated Container| Sandbox[Sandbox Client]
    
    Sandbox <-->|X11 / VNC Server| VNC[x11vnc + websockify] <-->|Websocket Stream| UI
    Sandbox <-->|Playwright / xdotool| Browser[Browser & OS Controls]

Module Breakdown Copy Section

🚀 Quick Start Copy Chapter

Prerequisites Copy Section

• Rust: 1.77.2 or later
• Node.js: 18.x or later

Install & Run Copy Section

1. Clone the Repository

   bash Copy

   git clone https://github.com/Onelevenvy/flock.git
   cd flock

2. Install Frontend Dependencies

   bash Copy

   cd flock-ui
   npm install

3. Start Development App

   bash Copy

   npm run tauri dev

4. Run Backend Tests

   bash Copy

   cargo test --workspace

📄 License Copy Chapter

Licensed under the Apache License, Version 2.0. See LICENSE for details.