Roo Code — AI Coding Team Inside VS Code

What it is

Roo Code is a free, open-source AI coding extension for VS Code. Originally forked from Cline, it has grown into a standalone product with custom agent modes, multi-model support, and MCP (Model Context Protocol) integration. It provides five built-in personas -- Code, Architect, Ask, Debug, and Custom -- each with scoped tool permissions so the right agent handles the right task.

It targets developers who want AI assistance directly in their editor without switching to a separate tool. Roo Code supports OpenAI, Anthropic Claude, local LLMs, and other providers through a unified configuration.

How it saves time or tokens

Each agent mode has scoped tool permissions, which means the Code agent can edit files but the Ask agent only reads them. This reduces unnecessary token consumption by preventing agents from running expensive tool calls when they only need to answer a question. The Custom mode lets you define exactly which tools an agent can access for specialized workflows.

MCP support means Roo Code can connect to external data sources (databases, APIs, documentation) without including all that context in the prompt, further reducing token usage.

How to use

1. Install from the VS Code marketplace:

code --install-extension RooVeterinaryInc.roo-cline

1. Open the Roo Code sidebar and configure your preferred AI provider (OpenAI, Anthropic, local LLM endpoint).

1. Select an agent mode (Code, Architect, Ask, Debug, or Custom) and start a conversation about your codebase.

Example

Using Roo Code's Code agent to add error handling:

[Code Mode] Add try-catch error handling to the database
connection function in src/db.ts. Log errors with the
existing Winston logger and return a typed error result
instead of throwing.

The Code agent reads the file, understands the existing patterns, generates the changes, and applies them with your approval.

Related on TokRepo

• Coding Tools -- AI-powered development tools for code generation and review
• Prompt Library -- Reusable prompts that work with AI coding assistants

Common pitfalls

• Selecting the wrong agent mode wastes tokens. Use Ask mode for questions about code and Code mode only when you want file edits. The Debug agent is optimized for error analysis with stack trace parsing.
• Large codebases can exceed context limits quickly. Use .roo/ignore to exclude vendor directories, build artifacts, and test fixtures from the agent's file access.
• Local LLM endpoints need sufficient model capability. Small models (under 7B parameters) struggle with multi-file editing tasks that Roo Code's Code agent performs.