HumanifyMe: Make AI sound like you

humanify_text · contextLabel: professional

What the agent wrote

"Hey John — I finished testing the flash script changes on the lab box today. It boots clean and the logs look good, so the PR is ready for review whenever you have time. Also, I'll be out tomorrow morning for an appointment but should be back by noon."

Real output from a real profile built on 3 writing samples. The em-dash is gone, "just letting you know" is this user's actual phrase, and "whenever you get a chance" is how they say it. Small shifts. That's the point, it's the small shifts that read as a person.

How it works

A voice fingerprint, not a filter.

"AI humanizers" rewrite toward a generic professional register, the opposite of sounding human. HumanifyMe builds a structured profile of how one specific person writes, and holds every rewrite to it.

Give it your writing

Paste 3-10 real emails or messages, point it at a folder of your writing, or import your ChatGPT / Claude history in one command. Everything is redacted (emails, phones, keys) before any model sees it.

humanifyme import chat export.zip
imported 384 samples

It learns your fingerprint

Not an opaque embedding, a readable, editable profile: sentence rhythm, formality, your actual phrases, the words you'd never use, how you apologize, how you disagree.

{
  "contractions": "always",
  "commonPhrases": ["sounds good"],
  "wordsToAvoid": ["delighted",
    "leverage", "circling back"]
}

Your agent uses it

Every MCP agent gets a humanify_text tool and bundled skills, so it reaches for your voice automatically, on PR descriptions, emails, anything you'll send as yourself.

you: draft a reply, then humanify it
agent: calling humanify_text…

Works where you work

One install. Every agent.

HumanifyMe is an MCP server, the open protocol your agents already speak. Build your profile once; every agent on your machine writes in your voice.

Claude Codeplugin or one-liner

Coworkplugin

Cursormcp.json

Claude Desktopconfig

Windsurfmcp config

Clinemcp config

Continueconfig.yaml

Zedsettings

Privacy

Your voice is yours. So is your data.

Trust is the entire product, so the architecture does the promising, not us.

Local-first by construction

Samples and your profile live in ~/.humanifyme/ on your machine. There is no HumanifyMe server. We could not read your data if we wanted to.

Redaction before anything leaves

Emails, phone numbers, addresses, card numbers, API keys, and tokens are masked before any text reaches an LLM, and restored in your copy afterward.

Your key, your provider, nothing else

The only outbound traffic is to the LLM provider you configure, Anthropic, OpenAI, Gemini, or a local Ollama that never touches the internet at all.

Auditable, deletable

Run humanifyme audit to see every outbound request, metadata only, never content. One command wipes everything, irrevocably.

$ humanifyme audit 2026-06-11T00:34:02 anthropic/v1/messages 6.2KB draft=261 profile=y ok 2026-06-11T00:29:00 anthropic/v1/messages 4.8KB draft=0 profile=n ok no other destinations. ever.

Install

Two minutes, one snippet.

Requires Node 22.5+. After installing, tell your agent "set up HumanifyMe", it walks you through provider, samples, and your first profile.

One command in your terminal:

claude mcp add humanifyme -- npx -y --package humanifyme@latest humanifyme-mcp

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "humanifyme": {
      "command": "npx",
      "args": ["-y", "--package", "humanifyme@latest", "humanifyme-mcp"]
    }
  }
}

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "humanifyme": {
      "command": "npx",
      "args": ["-y", "--package", "humanifyme@latest", "humanifyme-mcp"]
    }
  }
}

Same shape in your MCP settings JSON:

{
  "mcpServers": {
    "humanifyme": {
      "command": "npx",
      "args": ["-y", "--package", "humanifyme@latest", "humanifyme-mcp"]
    }
  }
}

Add to Zed settings.json:

{
  "context_servers": {
    "humanifyme": {
      "command": {
        "path": "npx",
        "args": ["-y", "--package", "humanifyme@latest", "humanifyme-mcp"]
      }
    }
  }
}

No agent needed, the full pipeline works from your terminal:

npx -y humanifyme setup
npx -y humanifyme provider set anthropic --api-key sk-ant-…
npx -y humanifyme import chat claude-export.zip --commit
npx -y humanifyme profile rebuild
echo "your stiff AI draft" | npx -y humanifyme rewrite

FAQ

Fair questions.

Is this an "AI detector bypass"?

No, and we don't make detector claims. HumanifyMe makes text sound like you because it's built from how you write, not from tricks aimed at fooling a classifier. Text that genuinely carries one person's voice tends to stop reading as AI to humans, which is the audience that matters.

What do you collect about me?

Nothing. There's no account, no telemetry, and no HumanifyMe backend. Your samples and profile stay in ~/.humanifyme/ on your machine. The only network traffic is to the LLM provider you configure, with your key.

What does it cost?

Free during beta. You bring your own API key, so your only cost is your provider's per-token price, typically well under a cent per rewrite. A managed plan (we handle the key) is planned later; bring-your-own-key stays free.

How many samples does it need?

It works from 3. It gets noticeably better with 10+, and best with a bulk import, your ChatGPT or Claude export gives it hundreds of examples of unguarded, real you in one command.

Can I see and edit what it learned?

Yes, that's the design. The profile is readable JSON with a plain-English summary (humanifyme profile show). Every field is editable, and edits stick: the profile is the source of truth, not the model.

Does it work with local models?

Yes. Point it at Ollama and nothing ever leaves your machine, not even redacted text.

Make AI sound
like you.