Etincel MCP

Self-hosted MCP server (stdio, no auth) — plus an optional hosted API — Etincel finds the AI tells in non-fiction prose deterministically, on your own machine: 20 tools that audit text for chatbot fingerprints, train brand voices from real writing samples, and measure drafts against a voice baseline, with a CLI, MCP server, and GitHub Action sharing one codebase (MIT).

Server type: Local (stdio) + optional hosted API
Auth: None for stdio; account only for the model-based `second_read` tool
Endpoint: npx etincel serve (local) · https://etincel.ai/api/mcp (hosted)
Tools: 20 (audit, voice training, style management, dictionaries)
Pricing: Free, open source (MIT); hosted second_read is a billed model call
Category: Content / Writing Quality
Built by: AIStoryHub (github.com/AIStoryHub/etincel, etincel.ai)

Why This Matters for Operators

Every operator shipping AI-drafted content faces the same risk: copy that reads synthetic destroys trust before the message lands. Etincel turns that risk into a deterministic check — audit_text scans for banned hype vocabulary, chatbot fingerprints, and structural patterns (uniform paragraph length, stacked transitions, em-dash overuse, rule-of-three compulsion) and returns a tiered verdict with each finding located by line. Because it is deterministic, the same text always gets the same verdict, which is what makes it safe to gate commits and CI runs on it.

The differentiator is the voice-training loop. train_style measures a real writer's habits — sentence length and variance, paragraph rhythm, contraction rate, em-dash and semicolon use, structural entropy — and persists them as a named voice. Drafts can then be checked against that baseline with check_voice_match, so brand voice stops being something reviewers eyeball and becomes something an agent can measure before content ships. The .etincelrc file makes the banned-words and house-style dictionary reviewable, versioned code, not a hidden vendor config.

Tools & Capabilities

Tool	Purpose
audit_text	Deterministic scan for AI-writing tells; tiered verdict (green/yellow/orange/red), scored findings with severity and location, plus a strengths counter-signal
train_style	Train a persistent brand voice from real writing samples (rhythm, contractions, punctuation habits, recurring phrasing)
check_voice_match	Measure a draft's rhythm against a trained voice baseline — verdict, match score, and per-dial drift notes
check_self_repetition	Detect a writer's own recurring openings and characteristic phrases across recent pieces
list_styles / get_style_guide	Enumerate presets and trained voices; fetch the full drafting guide for a style
create_style_from_dials	Build a voice from explicit dials (formality, warmth, directness) plus 8 mechanical dials — no samples needed
update_style / fork_style / delete_style / set_default_style	Rename or retune voices, fork public community styles, manage defaults
add_banned_word / remove_banned_word	Extend the banned-words list globally or per style
add_custom_word / remove_custom_word / list_dictionary	Maintain the allowed-words list (corporate dictionary: acronyms, house terms)
set_style_instructions / clear_style_instructions / get_style_instructions	Layer free-text drafting rules (required CTAs, forbidden topics) on top of any voice
second_read	One model call reading a draft like a careful human editor — hosted only, unscored, never a silent rewrite

Installation

claude mcp add etincel -- npx -y etincel serve

The CLI doubles as a linter for files in CI: npx etincel lint README.md, and the GitHub Action runs as AIStoryHub/etincel@main against pull requests.

Configuration

{
  "mcpServers": {
    "etincel-nonfiction": {
      "command": "npx",
      "args": ["etincel", "serve"]
    }
  }
}

No keys or accounts are needed for the local install — audit, training, and voice checks all run offline. The only hosted dependency is second_read, which requires an account at etincel.ai because it bills a model call; a .etincelrc at the repo root carries the shared team dictionary and style instructions (dictionary as code, reviewable and versioned).

Business Relevance

• Content operators get a pre-publish gate that flags AI tells with line references before copy reaches customers — tiered verdicts instead of binary detector scores.
• Brand and marketing teams can encode a real founder's voice from samples and check every draft against it, including drift warnings per dial.
• Engineering teams run the same checks in CI via the GitHub Action or npx etincel lint, so style gates survive the human reviewer's absence.
• Agencies can fork and reuse public community styles, then retrain them on client samples — voice work becomes a versioned asset, not a per-project retelling.

Integration with CorpusIQ

Etincel pairs with CorpusIQ as the quality gate on the content pipeline's output side. CorpusIQ connectors (GA4, Search Console, Stripe, Klaviyo) surface the performance data that tells you what to publish; Etincel checks the copy itself before it ships. The workflow composes cleanly: an agent drafts X posts, outreach emails, or docs pages, calls audit_text against the CorpusIQ house dictionary in .etincelrc (banned hype vocabulary, no em-dash rules, approved product terms), and only content that passes a green or yellow tier proceeds to publishing. Prose Coach (also in this catalog) covers the same problem as a no-auth remote service — Etincel is the self-hosted, CI-friendly complement when copy must stay in-house. For voice work, train_style can be fed real approved writing samples so future drafts are measured against the actual operator voice, not a generic tone prompt.

Limitations

• Brand new — submitted to mcp.so August 15, 2026; no track record or community adoption yet.
• Local-first: trained voices and self-repetition history live on the installing machine — a hosted/remote connection reports zero history today.
• second_read is the only model-based tool and requires a paid etincel.ai account; it fails with a clear explanation on pure stdio installs.
• check_voice_match measures rhythm and mechanics — it is explicitly not an authorship or AI-detection proof, and short inputs return low-confidence verdicts.
• Stdio transport means per-machine setup; there is no multi-user OAuth model like CorpusIQ's connectors.

See Also

• MCP Servers Index
• CorpusIQ Connectors