# Étincel

> A remote MCP server that gives Claude a trained writing voice for non-fiction prose. Connect it from Claude Code or Claude Desktop, pick a premade tone or train one from your own writing samples, and audit drafts for AI-writing tells.

Étincel is not an autocomplete or template engine. Trained voices are measured from real writing samples (sentence rhythm, contraction rate, em-dash habits, paragraph variance, recurring phrasing) or hand-tuned from ten style dials, then served to Claude as drafting context over the Model Context Protocol (MCP). The `audit_text` tool is a deterministic, rules-based scan for AI-writing tells — it flags issues with severity and specifics, never a silent rewrite.

Free to use, no paid plan. Sign-in via WorkOS AuthKit; trained voices are stored per-account in Postgres (Neon), not shared between users.

## Pages

- [Home](https://etincel.ai/): product overview, premade style presets, sign-in.
- [Connect via Claude Code](https://etincel.ai/connect): add Étincel as an HTTP MCP server with one terminal command.
- [Connect via Claude Desktop](https://etincel.ai/connect/claude-desktop): add Étincel as a custom connector from Settings.

## MCP endpoint

`https://etincel.ai/api/mcp` — Streamable HTTP transport, OAuth 2.1. Tools: `list_styles`, `get_style_guide`, `train_style`, `create_style_from_dials`, `set_default_style`, `audit_text`.

The same engine also ships as a local Claude Code / Claude Desktop plugin (stdio transport, no network calls, styles stored in `~/.etincel/`).