A prose linter and AI-slop detector. WSC finds AI tells — words, phrases, and sentence structures overrepresented in AI-generated text, each flag backed by a published corpus study. It also catches classic writing issues: weasel words, passive voice, duplicate words, long sentences, nominalizations, hedging, and filler adverbs. Available as a web editor, HTTP API, MCP server, CLI, and GitHub Action.
- Web Editor - Real-time highlighting with inline fix buttons for all 8 detectors
- HTTP API - POST text with optional config, retrieve structured JSON responses
- MCP Server (Remote) - Connect AI assistants via Streamable HTTP transport
- MCP Server (Local) - Stdio-based server via
wsc-mcpon npm - CLI - Check files from the command line via
wsc-lint - GitHub Action - Run checks in CI with
::warningannotations - Configurable - Customize detectors with
.wscrc.jsonfiles
WSC flags patterns that research on AI-generated text finds overrepresented, and cites a source for every flag. It does not, and cannot, prove authorship. Classifier-based detectors carry a documented false-accusation risk: a Stanford study found that seven of them misflagged 61% of essays written by non-native English speakers. WSC avoids that trap by design — every flag is a specific, explainable edit that improves the text no matter who, or what, wrote it.
| Detector | Items | Description |
|---|---|---|
| Weasel Words | 95 words/phrases | Vague terms like "very", "basically", "arguably", "numerous" |
| Passive Voice | 260 irregular verbs | Auxiliary verbs + past participles (regular -ed + irregular) |
| Duplicate Words | — | Adjacent repeated words across whitespace, case-insensitive |
| Long Sentences | threshold: 30 words | Sentences exceeding a configurable word count |
| Nominalizations | 245 word pairs | Nouns replaceable with verbs ("utilization" → "use") |
| Hedging | 100 phrases | Phrases that weaken assertions ("I think", "it seems") |
| Filler Adverbs | 139 words | Adverbs adding emphasis without substance ("totally", "utterly") |
| AI Tells | 98 words (+111 inflected forms) + 83 phrases + 12 structural patterns | Words, phrases, and sentence constructions overrepresented in AI-generated text (delve, rich tapestry, It's not just X — it's Y) |
Word lists sourced from Matt Might's shell scripts and expanded with additional entries. AI tells draw on published corpus studies: Kobak et al. 2025 (Science Advances), Juzek & Ward 2025 (COLING), Liang et al. 2024 (Stanford), and Reinhart et al. 2025 (PNAS). Wikipedia's editor-maintained Signs of AI writing catalogue and AI-detection vendor reports round out the sources.
Create a .wscrc.json to customize detectors. All tools (API, MCP, CLI) support it.
{
"$schema": "https://wsc.theserverless.dev/schema.json",
"ignore": ["CHANGELOG.md", "docs/vendor/**"],
"detectors": {
"weaselWords": {
"enabled": true,
"add": ["synergy", "leverage"],
"remove": ["very"]
},
"longSentences": { "maxWords": 25 },
"adverbs": { "enabled": false }
}
}Every field is optional. Missing fields use defaults. JSON Schema provides autocompletion in VS Code.
ignore is an array of glob patterns relative to the working directory. The GitHub Action uses the repository root. The CLI adds these patterns to its built-in **/node_modules/** and **/.git/** skips, together with the Action exclude input and the CLI --ignore flag.
Analyze text for writing style issues. Accepts optional config object.
curl -X POST https://wsc.theserverless.dev/api/check \
-H "Content-Type: application/json" \
-d '{"text":"The code was written very quickly."}'Response:
{
"summary": {
"total": 2,
"weaselWords": 1,
"passiveVoice": 1,
"duplicateWords": 0,
"longSentences": 0,
"nominalizations": 0,
"hedging": 0,
"adverbs": 0
},
"issues": {
"weaselWords": [{ "word": "very", "index": 21, "line": 1, "column": 22, "context": "..." }],
"passiveVoice": [{ "phrase": "was written", "index": 9, "line": 1, "column": 10, "context": "..." }],
"duplicateWords": [],
"longSentences": [],
"nominalizations": [],
"hedging": [],
"adverbs": []
},
"meta": { "characterCount": 34, "wordCount": 6, "sentenceCount": 1, "processingTimeMs": 2 }
}With config:
curl -X POST https://wsc.theserverless.dev/api/check \
-H "Content-Type: application/json" \
-d '{"text":"The code was written very quickly.", "config":{"detectors":{"weaselWords":{"enabled":false}}}}'Returns API documentation as JSON.
Returns the list of all 8 detectors with descriptions, configurability, and word counts.
Runs a smoke test with known text and returns {"status":"healthy"} or 503.
Limits: Max 100,000 characters per request. CORS enabled for all origins.
The Writing Style Checker is available as an MCP server, letting AI assistants check your writing directly.
| Tool | Description |
|---|---|
check_text |
Analyze text for all 8 writing style issues. Accepts optional config. |
fix_duplicates |
Remove duplicate adjacent words and return cleaned text |
list_word_lists |
Return info about all detector word lists |
check_file |
(Local only) Read a file from disk and analyze it. Auto-discovers .wscrc.json. |
Connect any MCP client to the hosted server - no installation required.
{
"mcpServers": {
"writing-style-checker": {
"type": "url",
"url": "https://wsc.theserverless.dev/mcp"
}
}
}Install via npm for local usage. Includes check_file for analyzing files on disk with auto-discovery of .wscrc.json.
npx wsc-mcpClaude Desktop / Claude Code config:
{
"mcpServers": {
"writing-style-checker": {
"command": "npx",
"args": ["wsc-mcp"]
}
}
}See the wsc-mcp npm page for full documentation.
Check files from the command line.
# Check all markdown files
npx wsc-lint check "**/*.md"
# Read from stdin
echo "The code was written very quickly." | npx wsc-lint check --stdin
# JSON output for scripting
npx wsc-lint check "**/*.md" --format json
# GitHub Actions annotations
npx wsc-lint check "**/*.md" --format github
# Create a config file
npx wsc-lint initSee the wsc-lint README for full documentation.
- uses: theserverlessdev/wsc@v1
with:
files: '**/*.md'
max-warnings: 20
exclude: |
CHANGELOG.md
vendor/**| Input | Default | Description |
|---|---|---|
files |
**/*.md |
Glob pattern for files to check |
config |
— | Path to .wscrc.json config file |
max-warnings |
unlimited | Max warnings before failing |
only-changed |
false |
On a pull request, check changed files that match files. If none match, skip the check. Other events check files. |
exclude |
— | Glob patterns to skip. Separate with a comma or a new line. Adds to config ignore. |
The web editor runs in your browser - we never send text to any server. The API and MCP endpoints only process text you explicitly send to them.
.
├── src/
│ ├── core/ # Shared detection engine
│ │ ├── detector.ts # 8 detection algorithms
│ │ ├── words.ts # Word/phrase lists (800+ entries)
│ │ ├── config.ts # Config types, merging, validation
│ │ ├── config-node.ts # Node-only: file loading, discovery
│ │ ├── analyzer.ts # Unified analyzeText() entry point
│ │ └── index.ts # Public API exports
│ ├── docs/ # Documentation content (Markdown files)
│ ├── mcp/
│ │ └── handler.ts # MCP JSON-RPC 2.0 handler
│ ├── lib/
│ │ ├── App.svelte # Main editor page component
│ │ ├── stores/theme.ts # Theme store (light/dark/system)
│ │ └── components/ # UI components (StatsBar, ConfigPanel, etc.)
│ ├── routes/
│ │ ├── +layout.svelte # Shared layout (header, nav, footer)
│ │ ├── api/check/+server.ts # HTTP API endpoint
│ │ ├── mcp/+server.ts # MCP endpoint
│ │ ├── health/+server.ts # Health check endpoint
│ │ ├── docs/+page.svelte # Documentation page
│ │ └── words/+page.svelte # Word library browser
│ └── styles/
│ └── main.scss # Global styles (light + dark themes)
├── mcp-server/ # Standalone stdio MCP server (npm: wsc-mcp)
├── cli/ # CLI tool (npm: wsc-lint)
├── action/ # GitHub Action (composite)
├── tests/ # 341 tests across 18 files
├── static/
│ ├── schema.json # JSON Schema for .wscrc.json
│ ├── llms.txt # AI/LLM discovery file
│ └── llms-full.txt # Detailed LLM context
├── wrangler.toml # Cloudflare Workers config
└── svelte.config.js # SvelteKit configuration
git clone https://github.com/theserverlessdev/wsc.git
cd wsc
npm install
npm run devVisit http://localhost:5173. The API is at /api/check, MCP at /mcp, health at /health.
| Command | Description |
|---|---|
npm run dev |
Start dev server |
npm run build |
Build for production |
npm run check |
Type check with svelte-check |
npm test |
Run all 341 tests |
npm run test:coverage |
Coverage report |
Deployed as a Cloudflare Worker at wsc.theserverless.dev.
npm run build
npx wrangler deploySee CONTRIBUTING.md for development setup, testing, and pull request guidelines.
For substantial changes, please open an issue first.
- Matt Might for the original shell scripts
- Built with SvelteKit and Svelte 5, deployed on Cloudflare Workers
- Logo made with DiffusionBee
