Etincel Nonfiction

What etincel-nonfiction does

The etincel-nonfiction MCP server combines two related capabilities: deterministic prose auditing and a voice layer for non-fiction drafting. Its audit identifies patterns that can make writing appear machine-generated, including uniform paragraph lengths, low sentence-length variation, mechanical register changes, and vocabulary or phrasing such as repeated high-frequency AI-associated terms. Findings include locations, severities, explanations, and suggested alternatives where applicable. The tool reports what it found rather than changing the source text.

The voice layer lets an author train a named style from their own writing. A trained style can be selected for individual requests or made the default. The repository also includes twelve presets covering emotional tones and use cases, including executive briefs, reflective essays, code comments, Slack messages, LinkedIn posts, website copy, and blog posts. Presets provide drafting context through formality, warmth, directness, sentence-rhythm, and voice descriptions; they are not fill-in templates.

How it works

The local server is designed for MCP hosts such as Claude Desktop and Claude Code, and can also be used with Cursor or other MCP clients. After installation, users make ordinary drafting requests and can ask the skill to write in a named voice or train a new one from sample text. The fork_style capability can turn a preset into a customized trained voice.

Auditing is deterministic and local. The underlying audit_text function powers both MCP usage and the command-line interface, so the same analysis can be applied in a conversation, a repository check, or continuous integration. The audit does not call a model, send text over the network, or decide who authored the document. It reports patterns and leaves edits to the author.

Setup and configuration

Run the local MCP server with npx etincel serve. A standard MCP configuration uses the npx command with etincel and serve as arguments. A local clone can instead be built with npm install && npm run build, then started with Node against dist/server.js.

The hosted deployment is available at https://etincel.ai/api/mcp over Streamable HTTP and uses account-based authentication. It is separate from this repository, which contains the local engine, CLI, and skill. MCP-aware clients that discover registry entries may find the server under ai.etincel/etincel-nonfiction.

The CLI accepts Markdown and Markdown-with-JSX files, automatically choosing the docs register for those extensions unless another register is supplied. It supports file globs, register selection, severity thresholds, and JSON output. A GitHub Action wraps the same lint command and can fail checks when findings reach the configured threshold.

Tools and capabilities

• Audit prose with audit_text and receive structured findings.
• Train named writing styles from user-provided samples.
• Select a style for one request or set a default style.
• Use twelve bundled voice and use-case presets.
• Customize a preset with fork_style.
• Lint documentation and other prose from the command line or GitHub Actions.

The repository reports measured AUC values by register rather than presenting the audit as authorship detection. Calibrated registers include docs, blog, memo, and essay; email and general are described as uncalibrated in the project documentation.