Roo Code and llama.cpp: local coding setup in VS Code

This Spanish-language tutorial sets up Roo Code in Visual Studio Code as a coding assistant connected to a local LLM through llama.cpp. The presenter selects an OpenAI-compatible provider, enters the server URL, enables streaming, and matches the extension's context window to the loaded model's configuration. Image support depends on loading a vision encoder; browser tools can still work without it.

The example is a to-do web app built with HTML, CSS, JavaScript and Bootstrap, with data stored in the browser. Architect mode produces a plan, while code mode implements it. Orchestrator mode delegates subtasks and returns their summaries to the main conversation. The tutorial also demonstrates file references and an agents.md summary for context in later conversations.

Permissions matter throughout the demonstration. The speaker warns that unrestricted command approval can affect files beyond the project. Git checkpoints provide a way to restore workspace changes, and the presenter uses one after an unclear browser-testing request leads to unwanted edits.

Playwright MCP exposes filtering and editing problems that code review missed. A narrower request to test buttons without modifying files gives clearer results. Brave Search then helps check a Python example for DeepInfra. That separate API exercise encounters credential and encoding errors. The demonstration shows local AI using external tools, but its search and hosted API steps require network access.