YOLO Agent
A VS Code coding agent powered by multiple AI providers. YOLO Agent gives you an interactive chat interface with tool-calling capabilities — file editing, terminal execution, diagnostics, sandboxed environments, MCP server integration, and more — all from the VS Code sidebar.
Hey! It's a human over here. So... I just wanted to chime in and let you know that this is just an experiment I'm trying to build in my free time. It's not polished, it might be buggy, and it may even break your code or mess with your files. Please use it with caution, and maybe even on a test project first. I don't want to be responsible for any lost work or corrupted files.
The whole idea behind this project is that most of the cheaper models DO eventually get to the right answer, but they need a lot of tries and some help from the user to get there. So I wanted to build a tool that makes it easy to have that back-and-forth interaction, where the model can ask for clarification, you can give it feedback, and it can iteratively improve its output until it's good enough. It uses a TODO system and asks the model to update the TODOs based on its own feedback, asking the model to clarify specifically what was implemented over and over until it gets it right. It's a bit like having a pair programming partner that doesn't get frustrated and can keep trying until it gets it right.
100% of the code in this repo was written by Claude Opus 4.6, GPT-5.3-Codex and Z.ai's GLM-5 using Kilo Code. I just stitched it together and added some glue code here and there. So if you find any issues or have suggestions, please keep in mind that it's not a polished product, but rather a fun experiment that I'm sharing with the world. Thanks for understanding!
Features
- Multi-provider support — Anthropic (Claude), OpenAI, OpenAI-compatible endpoints, Claude Code CLI, and Kilo Gateway.
- Multiple modes
- Sandboxed Smart To-Do (default) — Iterative plan → execute → verify loop with OS-level sandbox isolation.
- Smart To-Do — Same iterative loop without sandboxing.
- Code — Full autonomy with all tools.
- Architect — Design/planning-focused read-only strategy mode.
- Debug — Systematic troubleshooting and targeted fixes.
- Review — Read-focused code review and risk analysis.
- Orchestrator — Task decomposition and mode-switch coordination.
- Sandbox Orchestrator — Isolated development with git worktrees and OS-level restrictions.
- Ask — Chat only, no tool execution.
- Slash commands — Quick commands like
/debug, /review, /architect, /init, /summarize, and /newtask.
- Tool system — Read/write files, run terminal commands, fetch diagnostics, manage sandboxes.
- MCP integration — Connect to Model Context Protocol servers for additional tools.
- Context awareness — Automatically discovers skills, context files, and memory bank markdown.
- Multi-session — Run multiple agent sessions in the background and switch between them.
- Model selector — Switch models on the fly from the top bar.
- Kilo Auto Model routing —
kilo/auto now passes mode-aware routing hints to Kilo Gateway.
- Customizable settings — Kilo-style configurable chat/runtime options (autocomplete, queueing, context injection, compaction threshold, tool iteration limits) from both VS Code settings and the in-panel Settings UI.
Getting Started
Prerequisites
- VS Code 1.96+
- Node.js 18+
- npm (included with Node.js)
Building from Source
git clone https://github.com/FacuM/yolo-agent.git
cd yolo-agent
npm install
npm run compile
Installation
Option A — Development Mode
Open the cloned repository in VS Code and press F5 to launch the Extension Development Host with the extension loaded.
Option B — Install as a .vsix Package
Install the packaging tool:
npm install -g @vscode/vsce
Build the .vsix file:
vsce package
Install in VS Code:
code --install-extension yolo-agent-0.0.1.vsix
Or open VS Code → Extensions → ⋯ menu → Install from VSIX… and select the file.
- Open the YOLO Agent sidebar (activity bar icon).
- Click the ⚙︎ button → Providers tab → + Add Provider.
- Select an API kind (e.g.
anthropic), paste your API key, and pick a model.
Project Structure
src/
├── extension.ts # Activation entry point
├── config/ # Settings helpers
├── context/ # Workspace context scanner (skills, AGENTS.md)
├── mcp/ # MCP client, bridge, and config
├── modes/ # Mode presets, types, and manager
├── providers/ # LLM provider implementations and registry
├── sandbox/ # Sandbox manager (bubblewrap / git worktree)
├── sessions/ # Multi-session manager
├── tools/ # Tool definitions (file-ops, terminal, diagnostics, sandbox)
└── webview/ # Webview panel, HTML, JS, CSS
Development
npm run compile # One-shot build
npm run watch # Rebuild on change
npm run lint # ESLint
npm run test:e2e # VS Code integration tests
The build is handled by esbuild — see esbuild.js.
License
MIT
