humanpen-mcp
[License]
[MCP]
[Node]
English · 简体中文 · 日本語
Website · Pricing · Developer docs
Keywords: ai humanizer, mcp server, model context protocol, turnitin ai detection, reduce ai score, humanize ai text, bypass ai detection, docx ai humanizer, ai content rewriter, ai writing tool, claude mcp, cursor mcp, ithenticate ai report
Humanize what's flagged. Preserve the rest. An MCP server for HumanPen — a document-level AI humanizer that can humanize an entire document, rewrite user-selected passages, or automatically target flagged text from a Turnitin / iThenticate AI-detection report, editing .docx / .pptx files in place while preserving formatting, tables, images, citations, and formulas. Also converts citations between 12 styles, condenses to a word budget, and translates between 12 languages.
Features
- Selective rewriting by detection report — import a Turnitin / iThenticate AI Writing Report to pinpoint flagged passages; unflagged content is never touched
- Format in, format out — a DOCX comes back as a DOCX, a PPTX as a PPTX; formatting, tables, images, and formulas survive intact and the result is still editable
- Academic structure preserved — in-text citations, reference lists, footnotes, TOC fields, cross-references, figure numbering, equations, and special formatting are treated as protected objects
- No error injection — restructures meaning and syntax to change expression; never adds grammar mistakes, spelling errors, or awkward sentences as a detection strategy
- Full-length documents — no per-input word limit; a single file can be up to 100 MB, no splitting into text boxes
- Free to keep going — still flagged? Re-humanize for free with a fresh report until the AI rate falls to
* or 0%
- Word-count control (experimental) — set a min/max word range to keep the output within a target length
- Pay per rewrite — billed on words actually changed, not the whole document; failed and cancelled jobs cost nothing; credits never expire
Get a key
Sign up at https://humanpen.net and create a key at
https://humanpen.net/settings/api-keys. New accounts start with free credits,
enough to put a document through and see what comes back.
The key goes in an environment variable, never in a URL. URLs end up in server
logs, proxy logs, shell history and screenshots.
Install
Claude Code
-s user puts it in every project. The default scope is local, which
loads the server only in the directory you ran the command from — and looks
like a broken install the first time you open Claude Code somewhere else.
If your version rejects -e (reported
upstream), use the JSON
form:
OpenAI Codex
In ~/.codex/config.toml:
CodeBuddy / WorkBuddy
It also reads ${VAR} in its config, so the key can stay in your environment
instead of the file:
~/.codebuddy/.mcp.json for every project, <project>/.mcp.json for one.
Gemini CLI
It has gemini mcp add, but the argument order differs between versions — run
gemini mcp add --help and follow the usage line it prints. Pass the key with
-e HUMANPEN_API_KEY=... and the scope with -s user; the default is
project, which is only the directory you ran it in.
Claude Desktop
In claude_desktop_config.json. Use the absolute path to npx — run
which npx and paste the result: a desktop app is launched by the OS with a
minimal PATH, so the bare name that works in your terminal often is not found
here, and the only symptom is that the tools never appear.
Cursor / Windsurf / Cline
All three read the same shape — Cursor in .cursor/mcp.json, Windsurf in
~/.codeium/windsurf/mcp_config.json, Cline in its MCP settings panel:
OpenCode
In opencode.json — the key names differ slightly from everyone else's:
VS Code — keeps the key out of the config file
VS Code prompts once and stores the key in its secret store, so it never lands
in a file you might commit.
From source, or before the npm release lands
Then point your client at node /path/to/humanpen-mcp/dist/index.js instead of
npx -y humanpen-mcp.
Any MCP client works: this is a plain stdio server started by
npx -y humanpen-mcp with HUMANPEN_API_KEY in its environment.
Tools
Two things worth knowing
Jobs take minutes; tool calls do not. Each operation waits about 55 seconds
— enough for most documents — then returns a job_id with a note to call
check_job. The work continues on the server either way; nothing is lost by the
tool returning early.
ai_percent can be null, and that is usually good news. Turnitin prints
* instead of a number whenever AI writing comes in under 20% — it will not
quantify that band, because too much of it is false positives. So null means
"under 20%, and Turnitin will say no more", never "0%" and never "no result".
Questions people ask
Will this bring a Turnitin AI score down?
Usually under 20% in one pass with balanced — the threshold below which
Turnitin prints * instead of a number. If it misses, hand the result back with
the new report; only the passages still flagged get rewritten.
Does it work with iThenticate too?
Yes — pass either report. The format is read from the file.
Is my document sent to the model?
No. It uploads the file and answers with a path. A 40-page paper costs no tokens.
Privacy Policy
Documents you pass to a tool are uploaded over HTTPS to HumanPen's API
(api.humanpen.net) for processing; results are written back to your disk, and
processed files are kept server-side for about 7 days so check_job and the
free re-humanize pass can find them. Document contents never enter the model's
context. The full policy — what is collected, retention, and how to reach us —
is at https://humanpen.net/legal/privacy.
Development
selftest.mjs spawns the built server and talks JSON-RPC to it over stdio the
way a real client does — proving the protocol, the tool registrations, stdout
hygiene and one end-to-end job, not merely that the functions return. It needs a
live key and spends credits, so it is a pre-release check rather than a CI step.
Links
Apache-2.0
