Slop Cop - MCP Server

An MCP server that detects and fixes LLM prose patterns in text. It exposes tools to Claude Code, Claude Desktop, and any MCP-compatible host.

Built on the detection engine and browser app slop-cop by awnist that highlights these patterns in real time.

Setup

Claude Code

claude mcp add slop-cop -- npx -y @0ju1c3/slop-cop@latest

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "slop-cop": {
      "command": "npx",
      "args": ["-y", "@0ju1c3/slop-cop@latest"]
    }
  }
}

Restart Claude Desktop after saving.

Prerequisites

• Node.js 18+

Authentication

detect_slop runs entirely locally — no API key, no credentials.

fix_slop and detect_slop_full call the Anthropic API and require ANTHROPIC_API_KEY. Add it to your MCP server env config:

Claude Code:

claude mcp add slop-cop --env ANTHROPIC_API_KEY=sk-ant-... -- npx -y @0ju1c3/slop-cop@latest

Claude Desktop / other hosts:

{
  "mcpServers": {
    "slop-cop": {
      "command": "npx",
      "args": ["-y", "@0ju1c3/slop-cop@latest"],
      "env": { "ANTHROPIC_API_KEY": "sk-ant-..." }
    }
  }
}

Tools

detect_slop

Fast, instant detection using 36 regex and NLP rules. No API call, runs locally.

Use detect_slop on: "It is worth noting that this comprehensive solution
will fundamentally transform the landscape of modern development."

Returns a JSON array of violations:

[
  {
    "ruleId": "important-to-note",
    "startIndex": 0,
    "endIndex": 18,
    "matchedText": "It is worth noting",
    "ruleName": "\"It's Important to Note\"",
    "category": "rhetorical",
    "tip": "Delete the phrase and just say the thing."
  },
  {
    "ruleId": "elevated-register",
    "matchedText": "fundamentally transform",
    "suggestedChange": "",
    ...
  }
]

Best for: auditing prose, feeding violation data into a larger workflow, quick checks inside Claude Code sessions.

detect_slop_full

Runs all 36 client-side rules plus semantic analysis via the Anthropic API, catching the 12 patterns that require language understanding: sycophantic framing, throat-clearing, grandiose stakes, fractal summaries, and others.

Use detect_slop_full on: "And yes, I'll be honest — this challenge will
define the next era of computing for organizations that rise to meet it."

The LLM tier catches what regex can't:

• false-vulnerability — "And yes, I'll be honest" (performative candour)

• grandiose-stakes — "define the next era of computing" (inflated significance)

Best for: thorough audits of important prose before publishing.

fix_slop

Detects violations and fixes them in two phases:

Phase 1 — direct splices (instant, free): Violations with a known replacement are applied as exact character-position substitutions — "fosters" → "builds", "moving forward" → deleted, "utilize" → "use" — with no LLM call.

Phase 2 — targeted LLM rewrite (only if needed): Violations that require sentence restructuring (predicate adjectives, negation pivots) are rewritten paragraph-by-paragraph with rule-specific directives.

Use fix_slop on: "Ultimately, this transformative approach not only
streamlines workflows but also fosters a culture of innovation moving forward."

Returns: "This approach simplifies workflows and builds a culture of innovation."

Best for: cleaning up LLM-generated drafts before review, automating prose quality in content pipelines.

Use Cases

Passive quality gate in Claude Code

Add to your project's CLAUDE.md:

After generating any prose longer than 3 sentences, run detect_slop
on it and revise flagged violations before responding.

This makes slop detection automatic. Claude audits its own output before you see it.

Pre-publish audit

Before publishing a blog post, release notes, documentation, or other content generated with Claude's help:

Use detect_slop_full on the following draft: [paste text]

Content pipeline quality check

detect_slop returns structured JSON with ruleId, category, startIndex, endIndex, and suggestedChange. Any script or agent that generates text can pipe it through detect_slop and fail if violation density exceeds a threshold.

Inline editing

Ask Claude to run fix_slop on a specific paragraph you know sounds off. Phase 1 makes deterministic fixes; Phase 2 only fires if restructuring is needed.

Patterns Detected

Client-side — instant (36 rules)

Semantic — sentence-level (9 rules, via Claude Haiku)

Throat-Clearing Opener · Sycophantic Frame · Balanced Take · Unnecessary Elaboration · Empathy Performance · Pivot Paragraph · Grandiose Stakes · Historical Analogy Stack · False Vulnerability

Semantic — document-level (3 rules, via Claude Haiku)

Dead Metaphor · One-Point Dilution · Fractal Summaries

How It Works

Detection runs in two tiers:

Client-side (36 rules): Regex, structural analysis, and NLP via compromise. Runs entirely locally. It's instant, requires no API calls, and costs nothing.

Semantic (12 rules): API call to claude-haiku-4-5-20251001. Only runs when you call detect_slop_full or when fix_slop needs to restructure a sentence that a direct splice can't fix.

fix_slop applies all deterministic string splices first, at no cost, then calls the LLM only for violations that require sentence restructuring. Many texts are fully fixed in Phase 1 alone.

AI Detection Flow

detect_slop_full and fix_slop call an LLM in two steps:

MCP sampling first. The server calls sampling/createMessage to ask the MCP host to run a model on its behalf, as the MCP spec defines. Claude Code and Claude Desktop don't currently implement this; both return -32601: Method not found. The attempt is kept for forward compatibility with compliant hosts.

Direct Anthropic API as fallback. When sampling fails, the server calls the Anthropic API directly with claude-haiku-4-5-20251001 using ANTHROPIC_API_KEY. If the key is not set, the tool returns a clear error.

Building from Source

git clone https://github.com/0ju1c3/slop-cop
cd slop-cop
npm install
npm run build   # compiles TypeScript → build/

To use the local build instead of the npm package:

Claude Code:

claude mcp add slop-cop -- node /path/to/slop-cop/build/index.js

Claude Desktop:

{
  "mcpServers": {
    "slop-cop": {
      "command": "node",
      "args": ["/path/to/slop-cop/build/index.js"]
    }
  }
}

License

MIT — see LICENSE.

Detection engine derived from slop-cop by awnist, MIT licensed.

Pattern taxonomy based on LLM_PROSE_TELLS.md, Wikipedia: Signs of AI Writing, and tropes.md.