humanpen-mcp

Give your AI agent the ability to work on real documents. An MCP server for HumanPen. Point Claude, Codex, Cursor or any other MCP client at a .docx, .pptx or .pdf on disk, and it can lower the file's AI-detection score, convert its citations to another style, condense it to a word budget, or translate it — formatting, tables, images, citations and formulas intact.

English · 简体中文 · 日本語

Website · Pricing · Developer docs

claude mcp add humanpen -s user -e HUMANPEN_API_KEY=hp_your_key -- npx -y humanpen-mcp

"Here's my thesis and the Turnitin report — rewrite just the flagged parts, then put the references in IEEE style."

The agent reads the report, rewrites only the passages it marked, converts the citations, and hands back two file paths. It never had to read the thesis.

Why this instead of pasting text into the chat

It works on the file, not on a copy of the text. Paste a chapter into a chat and you get prose back — no headings, no tables, no figure numbering, no reference list, no equations. HumanPen edits the document itself and returns a document, so what comes out still opens in Word looking like what went in.

The document never enters the model's context. The server reads the file from disk, uploads it, and answers with the path it wrote. A 40-page paper costs you no tokens and is not copied into a transcript.

It can be guided by a detection report. Give it the Turnitin or iThenticate AI Writing PDF and it rewrites only the passages that report flagged, leaving everything else byte-identical. Rewriting a whole document to fix a quarter of it is how citations and meaning get damaged.

One tool call is one finished job. The server uploads, polls, downloads, and saves the result next to the source. No "here's a job id, remember to check it" loop for the model to lose track of.

Related MCP server: docx-forge-mcp

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 mcp add humanpen -s user -e HUMANPEN_API_KEY=hp_your_key -- npx -y humanpen-mcp

-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:

claude mcp add-json humanpen -s user '{"command":"npx","args":["-y","humanpen-mcp"],"env":{"HUMANPEN_API_KEY":"hp_your_key"}}'

In ~/.codex/config.toml:

[mcp_servers.humanpen]
command = "npx"
args = ["-y", "humanpen-mcp"]
env = { HUMANPEN_API_KEY = "hp_your_key" }

codebuddy mcp add --scope user humanpen -- npx -y humanpen-mcp

It also reads ${VAR} in its config, so the key can stay in your environment instead of the file:

{ "mcpServers": { "humanpen": {
  "command": "npx", "args": ["-y", "humanpen-mcp"],
  "env": { "HUMANPEN_API_KEY": "${HUMANPEN_API_KEY}" }
} } }

~/.codebuddy/.mcp.json for every project, <project>/.mcp.json for one.

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.

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.

{
  "mcpServers": {
    "humanpen": {
      "command": "npx",
      "args": ["-y", "humanpen-mcp"],
      "env": { "HUMANPEN_API_KEY": "hp_your_key" }
    }
  }
}

All three read the same shape — Cursor in .cursor/mcp.json, Windsurf in ~/.codeium/windsurf/mcp_config.json, Cline in its MCP settings panel:

{
  "mcpServers": {
    "humanpen": {
      "command": "npx",
      "args": ["-y", "humanpen-mcp"],
      "env": { "HUMANPEN_API_KEY": "hp_your_key" }
    }
  }
}

In opencode.json — the key names differ slightly from everyone else's:

{
  "mcp": {
    "humanpen": {
      "type": "local",
      "command": ["npx", "-y", "humanpen-mcp"],
      "environment": { "HUMANPEN_API_KEY": "hp_your_key" }
    }
  }
}

{
  "mcp": {
    "inputs": [
      { "type": "promptString", "id": "humanpenKey", "description": "HumanPen API key", "password": true }
    ],
    "servers": {
      "humanpen": {
        "command": "npx",
        "args": ["-y", "humanpen-mcp"],
        "env": { "HUMANPEN_API_KEY": "${input:humanpenKey}" }
      }
    }
  }
}

VS Code prompts once and stores the key in its secret store, so it never lands in a file you might commit.

git clone https://github.com/humanpen/humanpen-mcp
cd humanpen-mcp && npm install && npm run build

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.

Development

npm install
npm run build
HUMANPEN_API_KEY=hp_... node selftest.mjs sample.docx report.pdf

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

• API documentation · OpenAPI schema

• humanpen-skill — the same operations as an Agent Skill, if you would rather not run a server

• humanpen.net

Apache-2.0