Roo Code — AI Coding Agent with Custom Modes

What it is

Roo Code is a fork of Cline, the popular VS Code AI coding extension, that adds custom agent modes, boomerang orchestration, and multi-model routing. You can create specialized agents for different tasks: one mode for writing code, another for reviewing, another for architecture decisions. Each mode can use a different LLM and set of rules.

This tool is built for developers who want fine-grained control over how AI assists them. Instead of one generic coding assistant, you get a team of specialized agents.

How it saves time or tokens

By routing tasks to specialized modes, Roo Code avoids sending expensive context to the wrong model. A quick formatting fix can go to a fast, cheap model while a complex architecture question routes to a stronger one. The estimated token cost is around 4,000 tokens per session. Boomerang orchestration chains agents so the output of one becomes the input of the next, reducing manual copy-paste between tools.

How to use

1. Install Roo Code from the VS Code marketplace or build from source.
2. Configure your LLM API keys for the models you want to use.
3. Define custom modes with specific system prompts and model assignments.
4. Switch between modes during development using the command palette.

{
  "roo.modes": [
    {
      "name": "architect",
      "model": "claude-sonnet-4-20250514",
      "systemPrompt": "You are a senior software architect. Focus on design patterns, scalability, and maintainability."
    },
    {
      "name": "coder",
      "model": "claude-sonnet-4-20250514",
      "systemPrompt": "You are a senior developer. Write clean, tested code following project conventions."
    },
    {
      "name": "reviewer",
      "model": "claude-haiku-4-20250514",
      "systemPrompt": "You are a code reviewer. Find bugs, security issues, and style violations."
    }
  ]
}

Example

Boomerang orchestration chains modes:

1. [architect] Design the API schema for user authentication
   -> Output: API spec with endpoints and data models

2. [coder] Implement the API based on the architect output
   -> Output: Working code with tests

3. [reviewer] Review the implementation for security issues
   -> Output: Review comments and suggested fixes

Each step uses the optimal model for that task type.

Related on TokRepo

• AI coding tools — More AI-assisted development tools
• Multi-agent frameworks — Compare multi-agent orchestration approaches

Common pitfalls

• Too many custom modes creates decision paralysis. Start with 2-3 modes and add more only when you identify clear use cases.
• Boomerang orchestration requires well-defined handoff points. Vague output from one mode degrades the next mode's input quality.
• Different models have different context window sizes. Ensure your mode configuration accounts for the model's limits.
• Custom system prompts add token overhead to every request. Keep them concise and focused.
• Roo Code is a fork of Cline, so upstream Cline updates may not merge automatically. Check compatibility when updating.