Local scoring for AI agents

AI Detector MCP for Claude, Codex, Hermes Agent, OpenClaw

The Slop or Not MCP server lets Claude, Codex, Hermes Agent, OpenClaw, Cursor, and other agents call a local AI text detector, AI image detector, readability analyzer, and cleanup tool on your Mac.

MCP, short for Model Context Protocol, is a standard way for AI agents to call local tools. Unlike a cloud detector API, text and images sent here are processed by the signed Mac app, not a Slop or Not server.

Slop or Not MCP setup screen on Mac — Client setup snippets point at the bundled Mac app binary.

Local stdio server: Your MCP client starts the bundled slop binary and talks to it over stdio.
Six tools: Status, text detection, readability, cleanup, image detection, and raw image scoring.
Pro required: Tool calls return a Pro-required error when Pro is not active; the server keeps running.

Client setup

How do I add Slop or Not to an MCP client?

Install the Mac app, activate Pro, put slop on your PATH, then use the snippet for your client. Claude, Codex, Hermes Agent, OpenClaw, and Cursor all point at the same local server.

Claude CodeAdd Slop or Not as a user-scoped stdio server, restart Claude Code, then verify with /mcp.

claude mcp add --transport stdio --scope user SlopOrNot -- slop mcp

Claude DesktopUse this JSON shape in ~/Library/Application Support/Claude/claude_desktop_config.json, then restart Claude Desktop.

{
  "mcpServers": {
    "SlopOrNot": {
      "command": "slop",
      "args": ["mcp"]
    }
  }
}

CodexAdd the server to ~/.codex/config.toml, then restart Codex so it reads the new MCP server list.

[mcp_servers.SlopOrNot]
command = "slop"
args = ["mcp"]

Hermes AgentAdd Slop or Not as an MCP server in your Hermes Agent config, then restart Hermes Agent so it can call the local tools.

mcp_servers:
  SlopOrNot:
    command: "slop"
    args: ["mcp"]

OpenClawRegister Slop or Not with the OpenClaw MCP CLI, then restart OpenClaw if it was already running.

openclaw mcp set slopornot '{"command":"slop","args":["mcp"]}'

CursorAdd this to ~/.cursor/mcp.json for a global server, or .cursor/mcp.json at the project root for one project.

{
  "mcpServers": {
    "SlopOrNot": {
      "command": "slop",
      "args": ["mcp"]
    }
  }
}

Tool reference

What does the Slop or Not MCP server expose?

The server exposes six focused tools. Payload examples stay closed until you need them, so the page reads like a setup guide first and a reference second.

slop_status

Check app and Pro status

Confirms the app is installed, the binary can run, and Pro is active before the agent starts a workflow.

View payload and result

Tool input

{}

Result shape

{
  "pro": true,
  "version": "1.0.9"
}

detect_text

Detect AI text

Scores a passage with the on-device text detector and returns a local AI verdict, score, and readability metrics.

View payload and result

Tool input

{
  "text": "<text>",
  "include_readability": true,
  "language_code": "en"
}

Result shape

{
  "kind": "result",
  "verdict": "real",
  "score": 0.0,
  "language": "en",
  "sentence_count": 6,
  "generator": null,
  "readability": {
    "language": "en",
    "language_confidence": 0.9996,
    "scores": [
      { "kind": "fleschReadingEase", "value": 75.18 },
      { "kind": "fleschKincaidGradeLevel", "value": 5.51 }
    ],
    "stats": { "word_count": 66, "sentence_count": 6 },
    "warnings": [],
    "avg_words_per_sentence": 11,
    "word_count": 66,
    "sentence_count": 6
  }
}

analyze_readability

Analyze readability

Computes reading-level metrics without running AI detection.

View payload and result

Tool input

{
  "text": "<text>",
  "language_code": "en"
}

Result shape

{
  "language": "en",
  "language_confidence": 0.9996,
  "scores": [
    { "kind": "fleschReadingEase", "value": 88.54 },
    { "kind": "fleschKincaidGradeLevel", "value": 2.65 }
  ],
  "avg_words_per_sentence": 7,
  "sentence_count": 5,
  "word_count": 35,
  "warnings": []
}

clean_text

Clean text artifacts

Strips zero-width characters, homoglyphs, and fancy punctuation before the next detection pass.

View payload and result

Tool input

{
  "text": "<text>",
  "remove_invisibles": true,
  "remove_punctuation": true,
  "remove_homoglyphs": true
}

Result shape

{
  "cleaned_text": "<cleaned_text>",
  "language": "en",
  "removed_invisibles": 1,
  "punctuation_replacements": 1,
  "homoglyphs_replaced": 0,
  "british_substitutions": 0
}

detect_image

Detect AI images

Checks JPEG, PNG, HEIC, or WebP image bytes locally with C2PA and IPTC provenance reads and an on-device model fallback.

View payload and result

Tool input

{
  "image_base64": "<base64>",
  "recognize_text": true
}

Result shape

{
  "kind": "result",
  "verdict": "most_likely_ai_slop",
  "score": 0.80,
  "generator": null,
  "recognized_text": null,
  "recognized_sentence_count": null
}

score_image

Score AI images

Returns the raw OmniAID image score when an agent needs the model signal without the full image-detection response.

View payload and result

Tool input

{
  "image_base64": "<base64>"
}

Result shape

{
  "kind": "score",
  "score": 0.80
}

Verify

How do I verify the MCP server?

After restart, ask your agent to run slop_status. The expected result is a tool call that reports the local app and Pro state without an error.

{
  "pro": true,
  "version": "1.0.9"
}

Troubleshooting

What if my MCP client cannot find Slop or Not?

Some apps launch without your login shell PATH. In that case, point the MCP config directly at the binary inside the app bundle.

{
  "mcpServers": {
    "SlopOrNot": {
      "command": "/Applications/Slop Or Not - AI Fake Detector.app/Contents/MacOS/slop",
      "args": ["mcp"]
    }
  }
}

Local API

Can agents use this instead of a cloud AI detector API?

For agent workflows, yes. MCP gives Claude, Codex, Hermes Agent, OpenClaw, Cursor, and other clients a local tool interface instead of a hosted AI detector API. The client sends text or image data to the bundled Mac binary over stdio, and the check runs on your Mac.

Loop with agents

How does Slop or Not work with Agentic Humanizer?

Agentic Humanizer can call the MCP tools, score a baseline, clean mechanical artifacts, and re-score the draft locally. Optional voice matching belongs to the agent: your writing sample steers the rewrite, while Slop or Not handles the local measurement.

Slop or Not returns a probability verdict, not a guarantee. Results can vary with new AI models, short passages, and writing that was heavily edited by a human.

Read the MCP spec Claude Code MCP guide Agentic Humanizer on GitHub Read the MCP and CLI launch post Read the voice-matching guide Use the CLI instead