Étincel: Non-Fiction Writing Connector
___ ___ ___ ___ ___ ___
/\ \ /\ \ ___ /\__\ /\ \ /\ \ /\__\
/::\ \ \:\ \ /\ \ /::| | /::\ \ /::\ \ /:/ /
/:/\:\ \ \:\ \ \:\ \ /:|:| | /:/\:\ \ /:/\:\ \ /:/ /
/::\~\:\ \ /::\ \ /::\__\ /:/|:| |__ /:/ \:\ \ /::\~\:\ \ /:/ /
/:/\:\ \:\__\ /:/\:\__\ __/:/\/__/ /:/ |:| /\__\ /:/__/ \:\__\ /:/\:\ \:\__\ /:/__/
\:\~\:\ \/__/ /:/ \/__/ /\/:/ / \/__|:|/:/ / \:\ \ \/__/ \:\~\:\ \/__/ \:\ \
\:\ \:\__\ /:/ / \::/__/ |:/:/ / \:\ \ \:\ \:\__\ \:\ \
\:\ \/__/ \/__/ \:\__\ |::/ / \:\ \ \:\ \/__/ \:\ \
\:\__\ \/__/ /:/ / \:\__\ \:\__\ \:\__\
\/__/ \/__/ \/__/ \/__/ \/__/
Find the AI tells in your prose. Deterministically, locally, in CI.
npx etincel lint README.md
No install, no account, no model call, no network.
$ npx etincel lint drafts/announcement.md --register blog
✗ drafts/announcement.md RED 100/100 (8 findings, 169 words, register: blog)
Heavy AI styling. Multiple strong tells stacking up, worth a structural
rewrite, not a word-swap pass.
Whole-piece rhythm
medium uniform-paragraph-length
6 paragraphs, most running about the same length with little
variation between them. Vary paragraph length more.
medium low-burstiness
13 sentences averaging 13 words, with little variation in length
from one sentence to the next. Mix short sentences with long;
allow fragments.
medium mechanical-register-drift
Fragment rate and structural variety (sentence openers,
punctuation mix) sit off where blog prose typically lands. Allow
more sentence fragments and vary openers/punctuation more.
Vocabulary and phrasing
high comprehensive L3:C30 → thorough, complete
high cutting-edge L3:C109 → newest, latest
high leverage L3:C99 → use
high seamless L3:C150 → smooth, easy
high streamline L3:C70 → simplify, speed up
strengths specificity 11.4/1k · concrete:abstract 0.40 · burstiness 0.32
1 file audited, 1 at or above orange.
Étincel is two things. A deterministic audit that finds the patterns making prose read as machine-written, as a CLI, a GitHub Action, and an MCP tool. And a voice layer that trains Claude, Cursor, or any MCP client to draft in a style measured from your own writing, so the problem is smaller before you ever read the draft.
AI-drafted prose has a recognizable shape: uniform paragraphs, hedged authority, em dashes where a comma would do, closings that resolve too neatly. Étincel encodes the rules against that shape and, just as important, shows you what it found and why instead of quietly overwriting your voice. There is no model call anywhere in the audit path, and you decide what changes. You stay the author.
Does it work?
We measure it rather than assert it. Pooled AUC per register, measured by assay against labeled corpora with a fixed bootstrap seed, so any PR that drops these numbers fails CI:
Email is close to a coin flip. It's deliberately uncalibrated, and we track it anyway so an accidental regression doesn't go unnoticed. general is uncalibrated too. The calibrated registers are docs, blog, memo, and essay; use those if the number matters to you. See efficacy-baselines.json and src/data/SOURCES.md for method and provenance.
How this compares
vs. Vale + vale-ai-tells Vale is an excellent markup-aware linter and vale-ai-tells is a serious 78-rule package. They match tokens against patterns. Étincel measures shape: sentence-rhythm variance, paragraph-length uniformity, and repetition against your own previous drafts, the kind of statistical layer vale-ai-tells' own README says needs analysis beyond what token matching can do. Étincel also feeds a trained voice to the model before you draft, which no linter does. Running both is reasonable; they overlap less than you'd expect.
vs. skill files (stop-slop, avoid-ai-writing, no-slop) Those are instructions to a model, and they help. But a prompt can't verify it worked, can't produce an exit code, and can't gate a merge. Use one and run this in CI.
vs. humanizers (Undetectable.ai, QuillBot, StealthWriter) Those rewrite your text to evade detectors. Turnitin now has a dedicated AI-paraphrasing detection feature, expanded in 2025 to target humanizer output specifically. Étincel does the opposite job: it never modifies your text, and it would be a poor tool for fooling anyone.
vs. detectors (GPTZero, Pangram, Originality.ai, Copyleaks) Those answer "was a machine involved?" with a probability, over a whole document, as a verdict on authorship, in the cloud. Étincel answers "which specific patterns make this read as machine-written?" with locations and severities, locally, and never renders a verdict on who wrote something.
Install
Also listed on the official MCP registry as ai.etincel/etincel-nonfiction, so MCP-aware clients that auto-discover servers from there (VS Code's @mcp gallery, for one) find it without any of the config below.
Claude Code
/plugin marketplace add AIStoryHub/etincel
/plugin install etincel-nonfiction
Or from a local clone: /plugin marketplace add /path/to/etincel.
Claude Desktop / other MCP hosts
{
"mcpServers": {
"etincel-nonfiction": {
"command": "npx",
"args": ["etincel", "serve"]
}
}
}
From a local clone instead: npm install && npm run build, then point args at ["/path/to/etincel/dist/server.js"] with "command": "node".
Remote (hosted, multi-tenant)
A hosted version is also available at etincel.ai, exposing
the same tools over Streamable HTTP (https://etincel.ai/api/mcp) with
per-account auth instead of stdio. Point any MCP client at it directly:
{
"mcpServers": {
"etincel-nonfiction": {
"url": "https://etincel.ai/api/mcp"
}
}
}
The hosted server isn't part of this repo; this repo is the local/stdio engine, CLI, and skill that the hosted version is built on top of.
Using it
Once installed, just ask for what you'd normally ask for, like "draft an email to the team about the delay," "write a blog post about X," or "clean up this memo," inside Claude Code or Claude Desktop. The skill picks up automatically for non-fiction prose of meaningful length. To train your own voice:
Train a style called "me" from these three emails I wrote: [paste samples]
Then either name it per-request ("write this in my voice") or set it as default:
Set my default style to "me"
Style presets
Twelve premade presets ship out of the box: six emotional tones (Direct & Warm, Executive Brief, Reflective Essayist, Founder Memo, Plainspoken Analyst, Wry & Candid) plus six use-case presets (PR Review, Code Comment, Slack Message, LinkedIn Post, Website Copy, Blog Post). Each carries formality/warmth/directness dials plus a sentence-rhythm and voice description that gets fed to the model as drafting context, not a template that fills in blanks. The server reads these from src/data/presets.json. Fork any preset into a trained voice with fork_style to make it your own.
Command-line lint
audit_text is a pure function under the hood, so it also ships as a CLI, for linting prose outside a chat client (READMEs, docs, PR descriptions in CI):
npx etincel lint 'docs/**/*.md'
npx etincel lint README.md --register docs --threshold yellow
Exits non-zero if any matched file's tier is at or above --threshold (default orange). .md/.mdx files default to the docs register automatically (suppresses the Markdown-structure false positives, since a real heading isn't a chatbot tell); pass --register to override. Add --json for a machine-readable report. Run npx etincel lint --help for the full option list.
A GitHub Action wraps the same CLI (see action.yml, and .github/workflows/lint.yml in this repo for a working example):
- uses: AIStoryHub/etincel@main
with:
patterns: "docs/**/*.md README.md"
threshold: orange
Repo-local config: dictionary, instructions, and a shared team style
A team's rules don't have to live only in each person's local ~/.etincel/. Drop a .etincelrc (or .etincelrc.json / etincel.config.json) at the repo root and it's picked up automatically by the CLI and by the local (stdio) server, reviewable in code review and versioned instead of invisible and gone when someone leaves:
{
"bannedWords": ["Acme Cloud Platform"],
"allowedWords": ["leverage"],
"register": "docs",
"threshold": "orange",
"instructions": "Always include a one-line CTA at the end.",
"style": {
"name": "House Voice",
"dials": {
"formality": 6,
"warmth": 4,
"directness": 7,
"sentenceLength": 40,
"sentenceRhythmVariance": 50,
"paragraphVariance": 30,
"contractionUse": 20,
"emDashUse": 0,
"fragmentTolerance": 10,
"questionUse": 5,
"entropy": 60
}
}
}
bannedWords/allowedWordsmerge alongside whatever's in your account/style dictionary;register/thresholdact as repo-wide defaults that an explicit--register/--thresholdflag still overrides.instructionsis free text, folded intoget_style_guide'sinstructionsfor every style, not just the team one, ahead of your own account-level global instructions: the team-wide equivalent ofset_style_instructionswith nostyleId, but committed to the repo instead of living in one person's account.styledefines a shared "house voice" from dials, addressable everywhere asstyleId: "team"(get_style_guide, and once forked into a real trained voice withfork_style, everywhere else too) so a team has one already-tuned starting voice from day one instead of everyone hand-training or hand-tuning their own from scratch.list_stylesincludes it automatically when a.etincelrcin the current repo defines one.
The hosted server doesn't use any of this (it has no local repo to look in).
Sharing config across a team without committing it
.etincelrc is the versioned, code-reviewable layer above; the layer beneath it is ETINCEL_HOME, an environment variable that points the local (stdio) server and CLI at a directory to use instead of the default ~/.etincel/. Point every teammate's ETINCEL_HOME at the same shared, synced, or mounted directory (a repo-external path everyone's machine can read, e.g. something synced by your usual file-sharing setup) and trained voices, the default style, and account-level instructions/dictionaries are shared too, not just the .etincelrc-committed subset:
{
"mcpServers": {
"etincel-nonfiction": {
"command": "node",
"args": ["/path/to/etincel/dist/server.js"],
"env": { "ETINCEL_HOME": "/path/to/shared/etincel-home" }
}
}
}
Nothing else to export or import: pointing ETINCEL_HOME at the same directory is the sync, the same way it already is for a single person's ~/.etincel/.
What's in here
- An MCP server (
src/server.ts) exposing twenty tools:list_styles: premade tone presets, any voices you've trained, and (if a repo-local.etincelrcdefines one) a shared team styleget_style_guide: the drafting instructions for one styletrain_style: learn a voice from your own writing samples (sentence rhythm, contraction rate, em-dash habits, paragraph variance, recurring phrasing: measured, not guessed)create_style_from_dials: build a style from explicit formality/warmth/directness and mechanical dials instead of samplesupdate_style: rename a trained voice or adjust its dials in placefork_style: copy a preset's dials and guide into a new trained voice you can retrain or hand-tune, or fork another installer's style once they've published it publicly on the hosted gallery (addressed ashandle/slug, e.g.jpleblanc/blunt-memo, the same address shown on its public page atetincel.ai/v/handle/slug); a public-style fork also carries over its mechanical dials and any dictionary/instructions the source installer set specifically for that style (never their private, account-wide ones), and makes one network call toetincel.aito fetch it, while a preset fork never leaves this installdelete_style: permanently remove a trained voiceset_default_style: remember which style to use without repeating yourselfcheck_voice_match: compare a draft's measured rhythm against a trained voice's baseline. A rhythm/mechanics check, not an authorship or AI-detection check, and low-confidence on short inputcheck_self_repetition: compare a draft against a voice's own recent training samples for habits, not AI tells: the same opener, or a phrase, recurring across several past pieces ("you've opened this way in 4 of your last 6 pieces"). Local install only for nowaudit_text: a deterministic, rules-based scan for AI tells, returning a tier, specific findings with severity, and a strengths signal (specificity, concrete-vs-abstract ratio, sentence-rhythm variation) so fixes don't flatten the prose. Takes an optionalregister(email/blog/memo/essay/social/docs/general/personal, defaultgeneral) to calibrate strictness against the kind of text it is:docssuppresses Markdown-structure false positives (headings, bolded terms) and recalibrates rhythm/vocabulary detection against long-form reference prose instead of punchier short-form copy. is scaffolded (accepted, suppresses nothing extra yet) but not yet calibrated: no term suppressions, rhythm weight, or labeled corpus of its own until one is measured. Also takes an optional (details elicited from the user, never generated): checks how many actually made it into the draft, flagging below a quota of two used and at least one in a sentence that isn't proving a qualification
Trained voices, dictionaries, and your default style live locally in ~/.etincel/: nothing is sent anywhere. audit_text is plain deterministic code (string analysis + a curated corpus of AI-writing tells), not a model call. The one exception is forking a public style via fork_style, which fetches (never sends) that style's guide from etincel.ai's public gallery; forking a preset, or anything else in this list, still touches the network not at all.
Custom dictionaries
Beyond the built-in AI-tell corpus, you can maintain your own banned and "always allowed" word lists: just tell Claude (or any MCP client) things like "add [word] to my banned words list" or "add [word] to my custom words list, it's one of ours." Each list lives at a scope: global (applies everywhere, the default when no style is named) or a specific style id, whose list is merged on top of global when you audit against that style. list_dictionary shows what's saved for a scope, plus the effective merged list for a style. Editing the global list is already the way to keep a word in sync across every style: it's merged in automatically, live, every time audit_text or list_dictionary runs.
Development
npm install
npm test # run the engine/tools test suite (node:test via tsx)
npm run dev # run the MCP server over stdio via tsx, for local testing
npm run build # compile to dist/
The mcpscore badge audits the hosted remote server (etincel.ai/api/mcp), not the local/stdio engine in this repo; the two expose the same tools but run as separate deployments.
Status and limitations
Early, and honestly so.
- The audit corpus in
src/data/is a curated subset, not exhaustive. Seesrc/data/SOURCES.mdfor provenance and what isn't ported yet. The full corpus is at aistoryhub.co/corpus. email(AUC 0.540) andgeneral(0.748) are uncalibrated.socialhas no labeled corpus at all and isn't tracked.check_voice_matchis a rhythm-and-mechanics check, not authorship detection, and it's low-confidence on short input.check_self_repetitionis local-install only for now.- AI tells are a moving target: OpenAI patched ChatGPT's em-dash habit in November 2025. That's why efficacy is measured rather than claimed: when a tell stops being a tell, the AUC moves and the regression check catches it.