MCP Server
Use UndetectedGPT from inside Claude, ChatGPT, Cursor and other AI apps over the Model Context Protocol.
Overview
The Model Context Protocol (MCP) is an open standard that lets AI assistants call external tools. UndetectedGPT runs a remote MCP server that exposes the humanizer as a tool, so an assistant can rewrite its own drafts mid-conversation with no copying between tabs and no code.
The server speaks the Streamable HTTP transport and works with any client that supports remote MCP servers. For a guided, per-app setup, use the MCP server page in the dashboard.
Endpoint & authentication
POST https://www.undetectedgpt.ai/api/mcp
Authentication uses your regular API keys, in either of two ways:
Bearer header (preferred). For clients that can send custom headers, such as Claude Code, Cursor and n8n:
Authorization: Bearer ugpt_live_a1b2c3d4e5f6...
Keyed URL. For clients that only accept a URL, such as Claude.ai and ChatGPT custom connectors, append the key to the endpoint path:
https://www.undetectedgpt.ai/api/mcp/ugpt_live_a1b2c3d4e5f6...
The protocol handshake (initialize and tools/list) works without a key so connectors can be added and validated first; only tool calls require authentication.
The server implements the Streamable HTTP transport with JSON responses. The legacy SSE transport is not supported, so in clients that offer a transport choice (such as n8n), pick HTTP Streamable.
Tools
The server exposes two tools. Your assistant discovers them automatically and decides when to call them; you just ask in plain language.
humanize_text rewrites AI-generated text so it reads naturally and scores as human on AI detectors. It returns only the rewritten text and accepts the same options as POST /humanize:
text
string, required
The text to humanize. 5 words minimum; maximum set by your key (1000 by default).
tone
string
academic, balanced, conversational, formal or creative. Omit, or use balanced (equivalent), for the default style with no tone rewrite.
model
string
ghost-2 (default) or ghost-1.
ultra_stealth
boolean
Stronger restructuring for maximum detector evasion. Ignored when a rewrite tone is set.
markdown
boolean
Preserve markdown formatting. Defaults to true.
spelling
string
us (default), uk, uk-oxford, ca or au. English text only.
get_account_status reports the remaining word balance, the per-request word limit and the per-minute rate limit of the connected account, so the assistant can check whether a long job fits before starting it.
Connecting clients
The MCP server page walks through setup for Claude, ChatGPT, Claude Code, Cursor, Codex and n8n, with copy-ready snippets personalized to your key. As a taste, connecting Claude Code is a single command:
claude mcp add --transport http undetectedgpt \ https://www.undetectedgpt.ai/api/mcp \ --header "Authorization: Bearer YOUR_API_KEY"
Billing & limits
Billing
Identical to the API: one input word costs one word from your balance. Failed requests are refunded automatically.
Limits
Your key's per-request word limit and per-minute rate limit apply unchanged.
Usage
MCP calls appear on the Usage page alongside API calls and count toward the same totals.
Error handling
Tool failures are returned in-band as tool results the assistant can read and react to, rather than protocol errors. Each message states what went wrong and what to do next:
Missing or invalid key
The assistant is told how to authenticate and where to create a key.
Word limit exceeded
The assistant is told the limit and will split the text into chunks automatically.
Insufficient balance
The assistant reports the shortfall and points to the billing page.
Rate limited
The assistant is told to wait briefly before retrying.
Next steps
Was this page helpful?