GoHumanize Open Humanizer MCP server

An MCP (Model Context Protocol) server that lets AI assistants call the GoHumanize Open Humanizer: a small open model (Qwen3-4B fine-tune, Apache-2.0) that rewrites AI-styled English text into more natural human prose.

Tools:

• humanize_text rewrite a passage (50 to 400 words works best; longer texts are processed paragraph by paragraph, up to 1,500 words).
• about_open_humanizer what the model is and which endpoint is in use.

The model is educational and makes no claim about AI detectors. It is separate from the production models used by GoHumanize.ai.

Several rewrites, best one returned

On modern prose the model sometimes plays safe and hands the text back almost unchanged (about one try in six). Each call asks the endpoint for five rewrites (generated in parallel, so the wait is the same) and keeps the one that moved furthest from the input while staying a sensible length. Set OPEN_HUMANIZER_SAMPLES=1 for a single request.

Links

| Resource | Link | | --- | --- | | Project page and browser demo | gohumanize.ai/open-model | | GoHumanize (the product this research comes from) | gohumanize.ai | | Model weights and GGUF builds | gohumanize/gohumanize-open-humanizer | | Dataset, 3,257 pairs (CC-BY 4.0) | gohumanize/gohumanize-open-humanizer-dataset | | Code and full pipeline | GoHumanize-ai/gohumanize-open-humanizer | | Write-up: every step, service and result | docs/paper.md | | Archived release, citable DOI | 10.5281/zenodo.22843083 | | Python client and CLI | pypi.org/project/gohumanize-open-humanizer | | MCP server for AI assistants | npm · source | | Training runs, loss curves and config | Weights & Biases |

Use

{
  "mcpServers": {
    "gohumanize-open-humanizer": {
      "command": "npx",
      "args": ["-y", "gohumanize-open-humanizer-mcp"]
    }
  }
}

Which endpoint

The model runs wherever you point the server. Running it yourself needs no key and is the recommended setup:

ollama pull hf.co/gohumanize/gohumanize-open-humanizer:Q4_K_M
export OPEN_HUMANIZER_URL=http://localhost:11434/v1
export OPEN_HUMANIZER_MODEL=hf.co/gohumanize/gohumanize-open-humanizer:Q4_K_M

The hosted endpoint sleeps when idle. The first request after a quiet period waits for a GPU cold start, measured at one to two minutes; afterwards a rewrite takes a second or two. MCP clients apply their own timeout, often 60 seconds, so the first call through a client may fail even with a valid key and succeed on retry. Running the model locally avoids this entirely.

The endpoint the server falls back to is the one behind the browser demo on gohumanize.ai/open-model. It is rate-limited and requires OPEN_HUMANIZER_API_KEY, so it is not open for general use; to try the model without installing anything, use the demo on that page.

Any OpenAI-compatible server works:

"env": {
  "OPEN_HUMANIZER_URL": "http://localhost:11434/v1",
  "OPEN_HUMANIZER_MODEL": "gohumanize/open-humanizer"
}

OPEN_HUMANIZER_API_KEY sets a bearer token when the endpoint needs one.

Licence

Apache-2.0.