humanpen-mcp is a server based on the Model Context Protocol (MCP) that encapsulates HumanPen's document-level AI text humanization capabilities. It is aimed at users who need to process academic or professional documents and want to reduce the AI detection rate.
Main capabilities:
- Humanize an entire .docx or .pptx document without splitting text boxes; a single file can be up to 100 MB, with no per-input word limit.
- Rewrite paragraphs selected by the user, or automatically locate and rewrite only the flagged text based on a Turnitin / iThenticate report, without modifying unflagged content.
- Edit in place and preserve the original formatting: formatting, tables, images, formulas, in-text citations, reference lists, footnotes, table of contents fields, cross-references, figure and table numbers, equations, and special formatting are all treated as protected objects.
- Support conversion between 12 citation formats, compression to a target length based on a word budget, and translation between 12 languages.
- Does not use an error-injection strategy (no adding grammar or spelling errors, no creating awkward sentences); instead, it changes the expression by reorganizing semantics and syntax.
- Billing is based on the actual number of words rewritten; failed and canceled tasks are not charged, and credits do not expire. If it is still flagged, you can re-humanize for free with a new report until the AI rate drops to * or 0%.
- Provides an experimental word count control feature, allowing you to set a minimum/maximum word count range to constrain the output length.
How to use: At https://humanpen.net register, and at https://humanpen.net/settings/api-keys create an API Key. New accounts come with free credits, enough to complete a trial of one document. The API Key should be placed in an environment variable (never write it into a URL, to avoid leaking it into server logs, proxy logs, shell history, or screenshots). Run as a stdio server, started via npx: npx -y humanpen-mcp, and provide HUMANPEN_API_KEY in the environment. It supports any MCP client, including Claude Code, Codex, CodeBuddy, Gemini CLI, Claude Desktop, Cursor, Windsurf, Cline, opencode, VS Code, etc. Different clients differ slightly in configuration method and scope (local / user / project), so refer to their respective documentation. If the client version rejects the -e parameter, you can use a JSON-format configuration. You can also build it locally and then point the client to node /path/to/humanpen-mcp/dist/index.js.
Use cases: Suitable for users processing academic papers, reports, presentations, and other documents that need to preserve complex formatting and who want to reduce the AI detection rate, especially for scenarios where precise rewriting is needed after receiving a Turnitin / iThenticate flag.
Limitations and notes: Each operation tool call waits about 55 seconds, then returns a job_id, and you need to call check_job to continue checking; the job continues running on the server side and will not be lost. ai_percent may be null, which usually means Turnitin shows * (no specific value is shown when AI writing is below 20%), and it is not 0% or no result. Documents are uploaded via HTTPS to api.humanpen.net for processing, and the results are written back to the local disk. The processed files are retained on the server side for about 7 days so that check_job and free re-humanization can be used; document content does not enter the model context. For information about data collection and retention policies, please visit https://humanpen.net/legal/privacy。提供 selftest.mjs is used as a pre-release self-check, requires a valid Key, and consumes credits. There is also humanpen-skill, which provides the same operations in the form of an Agent Skill. For pricing, detailed prices, and developer documentation, please consult the official website.