Getting started

Install Caramelo, point it at an LLM provider, and run your first spec.

Install

From the Visual Studio Marketplace — search Caramelo in the Extensions sidebar and click Install.

Or from the command line:

code --install-extension fsilvaortiz.caramelo

You'll see a 🐱 icon in the activity bar — that's the Caramelo sidebar.

First run

1. Open the Caramelo sidebar. It has two panels: Providers (collapsed by default) and Workflow (visible).
2. Configure a provider. Expand Providers, pick a preset (Claude, OpenAI, Ollama, Copilot, …), enter your API key if needed, and click Save. The provider dot turns green when a one-token healthcheck succeeds.
3. Set up the constitution. From the Workflow panel, click Set up Constitution to begin. Either fill in principles by hand or use Generate with AI — describe your project in a sentence and the LLM proposes a starting set.
4. Create a spec. Click + New Spec, give it a name like user-auth, and a one-line description.
5. Run the phases. Each phase has a Generate button: Requirements → Design → Tasks. Approve each one to unlock the next. Documents stream into the editor as the LLM produces them.
6. Run the tasks. From the generated tasks.md, click Run Task on any line, or Run All Tasks from the editor's title bar. Tasks marked [P] run in parallel.

That's the full loop. The rest of this guide covers the things you can tune.

What you'll need

• VS Code 1.95 or later.
• An LLM provider — any of:
  • Cloud: Claude (Anthropic), OpenAI, Gemini, Groq, GitHub Copilot.
  • Local: Ollama or LM Studio.
  • Custom: any OpenAI-compatible endpoint (corporate proxies, self-hosted gateways).
• A workspace folder. Caramelo refuses to run agent tasks in untrusted workspaces.

Next

• Configure providers for the details of each preset, including custom auth headers for corporate proxies.
• Spec-driven flow walks through the four-phase ritual with screenshots.
• Agent loop & tools documents how /start-task actually edits your code.