4torm：多智能体协作框架，让 AI 智能体像长期团队成员一样常驻

Keep AI agents around like long-term team members, ready to collaborate when needed

4torm is a multi-agent collaboration framework. It lets you create persistent agents that can hold independent conversations, discuss problems as a team, hand work along workflows, or run tasks automatically on a schedule.

The same agent can be reused across different workspaces. Configure its model, prompt, tools, and skills once, then choose the collaboration mode that fits each task.

Five Workspaces

Workspace	Manages	Best suited for
Monsoon	Conversations	Ongoing conversations with one agent, tool use, and SubAgent delegation
Convection	Meeting rooms	Temporary multi-agent discussions with independent advice from a meeting assistant
Cyclone	Workshops	Organize fixed seats in a visual collaboration structure, with private chat, group chat, synchronous contact, and asynchronous dispatch
Tradewind	Workflows	Visual node-based workflows that pass tasks along connected paths
Tide	Automated tasks	Waking agents on a schedule for recurring checks, organization, and reporting

All five workspaces use the same agents. What changes is the form of collaboration—not the creation of another set of temporary roles.

Describe your goal in Monsoon, and an agent can use enabled framework collaboration tools to create meetings or workshops, build and start workflows, schedule tasks, and follow their results. You can also open the corresponding workspace to inspect progress and records or take part in the work.

Tools, Skills, and MCP

4torm extends agent capabilities in three ways:

Tools give agents executable operations for files, commands, networks, and other systems.
Skills provide reusable expertise and working methods through SKILL.md.
MCP connects external tool services over stdio, Streamable HTTP, and compatible SSE transports.

New agents have the framework's built-in tools and installed skills enabled by default. Custom tools and MCP tools must be selected explicitly.

Agents in Monsoon and Cyclone can also create new tools or skills at a human's request. Tool registration requires confirmation; completed skill directories are loaded automatically.

Quick Start

Requirements

Node.js 20.19+ (20.x) or 22.12+
npm 9+

Install Dependencies

git clone https://github.com/ccde141/4torm.git
cd 4torm

npm install
npm --prefix server install

If Windows PowerShell blocks npm.ps1, use npm.cmd in place of npm.

Start

Browser development mode:

npm run dev

Open http://localhost:5173.

Desktop development mode:

npm run electron:dev

Production build and server:

npm run build
npm run start:prod

Open https://localhost:3001.

After building, you can also run npm run electron:prod to open the desktop window.

First Use

Add a model provider, choose its actual provider type, and enable one or more models.
Create an agent from the Console.
Configure its model, role prompt, tools, and skills.
Start a conversation in Monsoon, or organize collaboration in another workspace.

Built-in presets cover generic OpenAI-compatible endpoints, dedicated cloud-provider configurations, and local runtimes such as llama.cpp, LM Studio, Ollama, and vLLM. Provider-specific parameters remain collapsed unless you need to tune them.

Documentation

Once the app is running, open the full documentation site from Documentation in the top navigation bar.

Topic	Documentation
Initial setup	Quick Start
Agents, sessions, and workspaces	Core Concepts
Creating and configuring agents	Agent Management
Five workspaces	Monsoon · Convection · Cyclone · Tradewind · Tide
Extending agent capabilities	Tools · Skills · MCP
Development and maintenance	Architecture · Data Layout · Security Boundaries

Preview the documentation site separately:

npm run docs:dev

Open http://localhost:5174.

Technology Stack

React 19, TypeScript, and Vite
Fastify 5
Electron
XY Flow
Server-Sent Events
Model Context Protocol