AI Slop Checker
About
Scores text and landing-page copy for AI-writing style tells; deterministic, offline, no LLM call and no network access.
Details
- Author
- parweb
- Categories
- Marketing, AI, Other
Jump to
Setup
Install AI Slop Checker in your MCP client (Claude Desktop, Cursor, Windsurf, and others).
Repository: https://github.com/parweb/mcp-ai-slop-checker
Follow the installation instructions in the repository README, then restart your MCP client.
An MCP server that tells your model when its own writing sounds like AI.
claude mcp add ai-slop-checker -- npx -y github:parweb/mcp-ai-slop-checker
Three tools, alldeterministic, local and offline: no LLM call, no API key, no network request, no telemetry. The same input always returns the same number, so you can put a score in a test and assert on it.
check_ai_slop(text) -> 0-100, 6 dimensions, named tells, fixes grade_landing_copy(headline, subhead, cta) -> 0-100, 5 dimensions, flags, rewrites get_slop_stats() -> benchmark stats from 239 real landing pages
Every "AI detector" is a probabilistic classifier that guesses at authorship and gets it wrong on both sides. This does the opposite and says so plainly: itcounts style tells— em-dash density,delve/tapestry/furthermorefrequency, "not only… but also" scaffolds, suspiciously even sentence lengths, missing specifics, over-parallel bullet lists — and hands back the raw counts that produced each sub-score.
That makes it useful in a loop an agent can actually close: write → score → see which count is high → fix that specific thing → re-score. A classifier's "87% likely AI" gives an agent nothing to act on."hype": 5does.
A score is a style measurement, not an authorship claim.stripe.comscores 61 and was obviously written by professionals. Low score meansreads generic, neverwas generated.
Listed in the officialMCP Registryasio.github.parweb/ai-slop-checker.
Installs straight from GitHub — not on npm yet, so use thegithub:spec:
claude mcp add ai-slop-checker -- npx -y github:parweb/mcp-ai-slop-checker
Or in any MCP client config (claude_desktop_config.json,.mcp.json, Cursor, etc.):
{ "mcpServers": { "ai-slop-checker": { "command": "npx", "args": ["-y", "github:parweb/mcp-ai-slop-checker"] } } }
Or install the self-contained MCPB bundle (dependencies included, no install step) from thev1.0.2 release—mcp-ai-slop-checker.mcpb, SHA-2566b13eb6d19be99553ab4551c7b6f9fc159a0db854c20718c611bfa0cc30f43f8. Rebuild it yourself and compare:./scripts/build-mcpb.sh.
git clone https://github.com/parweb/mcp-ai-slop-checker cd mcp-ai-slop-checker && npm install && npm test # then point your client at: node /abs/path/mcp-ai-slop-checker/src/index.js
Node >= 18. One runtime dependency (@modelcontextprotocol/sdk) pluszod.
Scores prose 0-100, where100 reads human. Six dimensions: LLM-word density (30), em-dash density (20), formulaic structures (15), sentence rhythm (15), specificity (10), list perfection (10). ~200+ characters gives a reliable read.
Real output, trimmed to the parts that matter:
// input: a 74-word paragraph of "In today's fast-paced world… delve… Moreover… seamless…" { "score": 34, "verdict": "This sounds AI-generated.", "words": 74, "dimensions": [ { "key": "LLM-word density", "max": 30, "score": 0, "notes": { "phrases": 5, "words": 14 } }, { "key": "Em-dash density", "max": 20, "score": 8, "notes": { "dashes": 1 } }, { "key": "Formulaic structures", "max": 15, "score": 10, "notes": { "hits": 1, "triads": 1 } }, { "key": "Sentence rhythm", "max": 15, "score": 6, "notes": { "sentences": 5, "cv": 0.2 } }, { "key": "Specificity", "max": 10, "score": 0, "notes": { "number": false, "propers": 0 } }, { "key": "List perfection", "max": 10, "score": 10, "notes": { "bullets": 0, "bold": 0 } } ], "flags": ["llmwords", "emdash", "formulaic", "uniform", "nospec"], "fixes": [ { "title": "Cut the LLM words", "detail": "Found 19 (\"delve/tapestry/furthermore/it's important to note\"…). Each one is a known model tell. Replace with the plain word you'd say out loud." }, { "title": "Vary sentence length", "detail": "Your sentences are suspiciously even (5 sentences, low variance). Humans write long, then short. Like this." } ] }
The hand-written paragraph intest/engine.test.js— same subject, same rough length — scores92, "Reads human."
grade_landing_copy(headline, subhead, cta)
Scores a hero block 0-100 across Anti-hype (25), Specificity (25), Clarity (25), Headline shape (13), CTA (12).subheadandctaare optional, but an empty CTA scores 0 on that dimension.
Three exclusions are worth knowing, because each one was a measured false positive rather than a preference:a digit that is part of a name, a version, a year or a list index is not a quantified claim(Auth0,Framer 3.0,B2C,© 2026);an arrow or a check mark is not an emoji(Get started →was losing 4 points for a button glyph); andan acronym is not shouting—SQL,MCP,CLI,APIno longer count as ALL-CAPS. Byte-for-byte the same rules as the browser grader inparweb/landing-copy-graderand the live one; verified identical on all 239 corpus pages.
// headline: "Revolutionize your workflow with our seamless, cutting-edge platform" // subhead: "Unlock powerful solutions that transform your business" // cta: "Learn more" { "score": 32, "verdict": "This reads AI-generated.", "dimensions": [ { "key": "Anti-hype", "max": 25, "score": 0, "notes": { "hype": 5, "exclamations": 0, "emoji": 0, "allcaps": 0 } }, { "key": "Specificity", "max": 25, "score": 8, "notes": { "number": false } }, { "key": "Clarity", "max": 25, "score": 7, "notes": { "filler": 3 } }, { "key": "Headline shape", "max": 13, "score": 13, "notes": { "words": 8 } }, { "key": "CTA", "max": 12, "score": 4, "notes": { "weak": true, "empty": false } } ], "flags": ["hype", "filler", "weakcta", "nonum"], "fixes": [ { "title": "Cut the hype words", "detail": "Found 5 (\"revolutionize/unlock/seamless/leverage\"…). Replace each with a plain, concrete verb." }, { "title": "Add one number", "detail": "No concrete figure anywhere. …82% of the 239 pages in our dataset fail this one." }, { "title": "Rewrite the CTA", "detail": "\"Learn more\" is generic. Use an action + outcome…" } ] }
Fix all four and the same offer scores100:
headline: "Cut invoice time from 3 days to 20 minutes" subhead: "Turn your spreadsheet into a client-ready invoice, no template hunting." cta: "Start your first invoice" -> { "score": 100, "verdict": "Reads human & sharp.", "flags": [] }
Both numbers are asserted intest/engine.test.js, so they can't silently drift.
Without a baseline, "your copy scored 74" is meaningless. This returns the reference distribution so the model can say"that's below the median of 239 real landing pages."
These are the figures of thedeposited corpus, scored withstatic-fetch-regex-v1. Three rules were tightened on 2026-07-25 — a digit inside a name/version/year is not a claim, an arrow is not an emoji, an acronym is not shouting — andgrade_landing_copyapplies them, soa page re-scored today can differ from its row in this table. The corpus deliberately keeps its original scoring: it is an archived object with a DOI, not a live view.
The most common tell is not the em-dash and not "delve" — it's the absence of a number.Four landing pages in five make a claim with zero quantity attached to it.
Full CSV with the extracted hero text of every page, methodology and the exclusion list:landing-copy-grader/data/landing-pages-scores.csv.node scripts/verify-dataset.jsin that repo re-scores all 239 rows offline and fails on any disagreement — the table above is pinned to its output bytest/engine.test.js.
Correction, 2026-07-25.These counts were wrong in v1.0.0 and are fixed onmain. The CSV they were computed from stored only the first three flags per row, so every page with four or more tells lost one:nonumread194 / 81%instead of195 / 82%, andcaps,shorthl,longhlandemojiwere low too. Scores, median, mean and the perfect-100 list were never affected.If you saw 194 / 81% from us anywhere, 195 / 82% is the correct figure.
18 tests: the scoring engines against published fixtures, plus 6 that spawn the real server over stdio and drive it through an actual MCP client (listTools, threecallToolround-trips, optional-argument handling, and a validation error that must not kill the process).
- Does this sound AI?—check_ai_slopin the browser
- Landing-page leaderboard— all 239 pages, scored, with the hero text
- parweb/landing-copy-grader— the single-file browser grader and the dataset
First published 2026-07-25.Small and young — stated plainly so you can judge it.
- Stable:the three tool signatures, the JSON shape they return, and the two scoring engines. Their outputs are asserted in the test suite, so a change that moves a score fails CI rather than surprising you.
- Opinionated and expected to change:the word lists. English only.
- Known gap:thev1.0.0bundle ships wrong benchmark numbers andv1.0.0/v1.0.1both ship the pre-correction scoring rules. Usev1.0.2, or thenpx github:install, which tracksmain.
Issues and PRs welcome, particularly on the word lists — "this term is wrong, here's a counter-example" is a reproducible bug report against a deterministic scorer, which is most of the point of building it this way.
- This counts style tells. It does not detect authorship, and nothing reliably does.
- The word lists are opinionated and English-only. They are plain arrays at the top ofsrc/slop.jsandsrc/copy.js— read them, disagree, fork.
- Scores are comparable over time only because nothing here is stochastic. That's the whole point.
- Built and maintained by an autonomous agent org. The code, the dataset and these numbers are real and reproducible; runnpm testand check.
Free MCP that drives an audit of your marketing. Your AI connects, adsOS digs through your ads, email and site, and hands back a growth plan you can run today.
A shared campaign canvas for you and your AI agent: briefed by your brand rules, gated by your approval.
Run your Linkedin account from claude or chatgpt
Chat with any brand's Meta (Facebook/Instagram) ads inside Claude — research a competitor's ad library, surface their longest-running winners, extract hooks/formats, and clone winning ads for your own brand.
Market & competitive intelligence for Claude & AI assistants: trends, alerts, keywords & more.
Intelligent Growth: Product Marketing MCP Server
Run positioning, messaging, competitive analysis and launch planning from Claude or ChatGPT with a hosted product marketing MCP server.
Outbound sales agent for Claude Code: per-prospect research, email writing, Gmail sending, reply tracking, and structured rejection feedback.
Manage mailtani email marketing from Claude, Cursor, or any MCP-compatible AI agent — contacts, lists, campaigns, segments, and event tracking.
Checks whether ChatGPT, Perplexity, and Gemini cite your brand, callable from any MCP client.
Google Ads reporting and campaign management for Claude; everything it creates starts paused, nothing spends until you turn it on.
Sign in to leave a review
Use Google, GitHub, or an email account so ratings stay tied to real people.
No reviews posted yet.
