# Dazzle

> Cloud stages for AI agents and live streaming.
> https://dazzle.fm

Dazzle gives you cloud stages — isolated browser environments that render and broadcast your content. Two tiers: GPU (NVIDIA RTX, hardware WebGL, 30 FPS shaders) and CPU (software OpenGL, lighter content). All at 1280x720 @ 30 fps.

## Docs

- [Full LLM Reference](https://dazzle.fm/llms-full.txt): Complete getting started, CLI reference, content guide — all in one file
- [Content Authoring Guide](https://dazzle.fm/guide.md): Detailed content guide with GPU vs CPU tiers, performance tips, design guidance
- [CLI Reference](https://dazzle.fm/cli-reference.txt): Full command listing
- [Stream Examples](https://github.com/dazzle-labs/stream-examples): Open-source example stages

## Quick Start

```bash
curl -sSL https://dazzle.fm/install.sh | sh     # install (Win: irm dazzle.fm/install.ps1 | iex)
dazzle login                                     # authenticate (or DAZZLE_API_KEY for CI)
dazzle s new my-stage && dazzle s up             # create + activate
dazzle s sync ./my-app --watch                   # push content (auto-refresh on change)
dazzle s ss -o preview.png                       # screenshot to verify
dazzle dest attach my-twitch                     # stream to Twitch/Kick/Restream/custom RTMP
```

## MCP Support

`dazzle mcp` starts a local MCP server on stdin/stdout — works with Claude Desktop, Claude Code, VS Code, Cursor, and any MCP client:

```json
{ "mcpServers": { "dazzle": { "command": "dazzle", "args": ["mcp"] } } }
```

**Tools:** `cli`, `screenshot`, `write_file`, `read_file`, `edit_file`, `list_files`, `sync`, `guide`.

Workspace tools (`write_file`, `read_file`, `edit_file`, `list_files`, `sync`) operate on `~/.dazzle/stages/<stage-id>/` on the host. This bridges sandboxed environments (e.g. Claude Desktop) where the agent's bash can't share files with the CLI process.

**Limitations vs full CLI access:** No shell/exec — can't run build tools (npm, npx, tailwind, etc.) in the workspace. Content must be pre-built HTML/CSS/JS. Agents with direct filesystem + shell access (e.g. Claude Code) should use `dazzle stage sync` directly instead.

## Notes

- Stages are **public during beta** — viewers see rendered video, not source code
- Content sync requires `index.html` entry point; `--watch` for live dev
- `localStorage` persists across restarts; live events via `dazzle s ev e <name> '<json>'`
- Relaxed CORS — stages can fetch any external API without cross-origin errors