humanizer-mcp

An MCP (Model Context Protocol) server that measures AI-detection risk in a piece of text and tells you — line by line — what to change. Works with Claude Code, Claude Desktop, and any MCP-compatible client.

Just want to use it? Go to humanizer.analyticadss.com — copy the URL, paste it into Claude's Connectors, done. The rest of this README is for developers and people self-hosting.

Rather than running your prose through a black-box "humanizer," this server analyzes it against known detection signals (vocabulary, burstiness, contraction usage, paragraph uniformity, em dashes, first-person voice) and returns a structured report with a 0–100 risk score and a concrete rewrite plan. The actual rewriting is left to the LLM that's driving the conversation — which is the point: a planner, not a laundering service.

Tools

Related MCP server: AI Content Detector MCP Server

Three ways to use it

Sharing with non-technical users? Two PDFs in share/ — email either one, no further explanation needed:

• share/humanizer-mcp-friends-guide.pdf — 2 pages. Truly non-technical: download a zip, upload it to Claude, done. No install.

• share/humanizer-mcp-boss-guide.pdf — 3 pages. Semi-technical Windows users who can install an app and edit a config file. Sets up the full MCP server with Claude Desktop.

Path A — add as a Custom Connector (zero install)

Works in claude.ai (web), Claude Desktop, and Claude for Chrome — all four surfaces share the connector list once you're signed in. Available on every plan including Free (Free is limited to one custom connector).

A hosted reference instance is up — feel free to use it for casual evaluation:

https://humanizer-api.analyticadss.com/mcp

For production / privacy-sensitive use, deploy your own with the included Dockerfile (see Hosting below — Fly.io takes ~3 minutes). The hosted instance is on a free Fly tier with no SLA, no support, and no privacy guarantees — your text passes through it.

To add it to your Claude:

1. Open Claude → Settings → Connectors.

2. Click Add custom connector.

3. Paste the URL above (or your own hosted instance's /mcp URL).

4. Save. The five humanizer_* tools become available in any chat.

That's the whole install for non-technical users — they never touch a terminal.

Path B — install locally (Claude Code / Desktop with stdio)

# Claude Code, one line
claude mcp add humanizer -- uvx humanizer-mcp

For Claude Desktop with a local stdio server, add this to claude_desktop_config.json:

{
  "mcpServers": {
    "humanizer": {
      "command": "uvx",
      "args": ["humanizer-mcp"]
    }
  }
}

Config location:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• Linux: ~/.config/Claude/claude_desktop_config.json

Other ways to launch the local binary if you don't want uvx:

pip install humanizer-mcp && humanizer-mcp     # pip
npx humanizer-mcp                              # npm launcher (delegates to uvx/pipx/python3)

Try it with the MCP Inspector

npx @modelcontextprotocol/inspector uvx humanizer-mcp

Hosting

To create the URL in Path A, deploy the included Dockerfile. The repo ships with a Render Blueprint and a Fly config:

Render — easiest, free tier, auto-deploys from the GitHub repo:

Fly.io — always-on free tier:

fly launch --copy-config --name humanizer-mcp
fly deploy

Anywhere else — the Dockerfile reads PORT from the environment and binds to 0.0.0.0, so it runs on Railway, Heroku, Cloud Run, ECS, or your own box:

docker build -t humanizer-mcp .
docker run -p 8000:8000 humanizer-mcp

Cloudflare Tunnel from your laptop — zero hosting cost, only up while your machine is on:

brew install cloudflared
pipx install humanizer-mcp        # or: pip install humanizer-mcp
humanizer-mcp --http --port 8000 &
cloudflared tunnel --url http://localhost:8000
# copy the trycloudflare.com URL it prints

The MCP endpoint is at /mcp (streamable HTTP). The server is stateless and unauthenticated — anyone with the URL can call the tools, but there are no secrets and no destructive operations to abuse.

Run the HTTP server locally

humanizer-mcp --http --port 8000
# point a client at http://127.0.0.1:8000/mcp

Verify it works

Once installed by any path, in any Claude chat ask:

"What humanizer tools do you have available?"

Claude should list five: humanizer_analyze_ai_tells, humanizer_quick_vocab_scan, humanizer_get_rewrite_instructions, humanizer_compare_before_after, humanizer_get_banned_words.

Then try the canonical test:

"Score this for AI tells: 'In today's rapidly evolving digital landscape, it's important to note that businesses must leverage cutting-edge solutions to navigate the multifaceted challenges they face.'"

You should get a score in the HIGH bucket (≥ 60), the signals that fired, and a line-by-line fix list.

Troubleshooting

Example prompts

With the server connected to Claude, you can say things like:

• "Analyze this blog post for AI tells and tell me what to change."

• "Run a quick vocab scan on this paragraph."

• "Give me rewrite instructions for this academic abstract — keep it formal but fix the burstiness."

• "Compare these two drafts. Did my edit actually lower the detection risk?"

Claude picks the right tool automatically.

How the risk score works

The 0–100 score combines eight signals:

1. AI vocabulary hits — words statistically overrepresented in LLM output (delve, crucial, leverage, myriad, …).

2. AI phrase hits — cliché structural tells (it's important to note, in the ever-evolving, at the end of the day, …).

3. Burstiness — coefficient of variation of sentence lengths. AI writing clusters around a single length; humans mix short fragments and long digressions.

4. Contractions — expanded forms (it is, do not) read as AI-formal; contractions read as conversational.

5. Paragraph uniformity — AI tends to produce paragraphs of similar size.

6. Rhetorical questions — near-absent in AI prose above 200 words.

7. First-person voice — AI avoids I, we, my, our unless prompted.

8. Em dashes — a ChatGPT signature; heavy use is a strong signal.

Each signal adds to the score independently; the total is clamped to 100 and bucketed into LOW (≤ 20), MEDIUM (21–50), or HIGH (51+).

Development

git clone https://github.com/aousabdo/humanizer-mcp
cd humanizer-mcp
pip install -e ".[dev]"
pytest

See CONTRIBUTING.md for more.

License

MIT — see LICENSE.