Use Stealth Agent from MCP clients

StealthGPT ships a remote MCP (Model Context Protocol) server, so AI agents and MCP-aware clients can generate and humanize content as a tool — without you writing any integration code. This is the agent / tool invocation workflow pattern: the model decides when to call the tools and you stay inside your editor or chat.

One server, many clients. The same MCP server works with Cursor, Claude, ChatGPT, and any MCP-compatible runtime. You don’t install a per-client plugin.

Server endpoint

https://stealthgpt.ai/api/mcp/mcp

Transport: Streamable HTTP (the modern MCP transport). A legacy SSE endpoint is also available at /api/mcp/sse.

Authentication

The server supports two ways to authenticate, and both resolve to the same StealthGPT account (identical billing):

OAuth (one-click) — used by claude.ai (web) and other clients with built-in custom-connector support. You sign in with your StealthGPT account and approve access; there’s no token to copy. See Connect in claude.ai (web).
API token (bearer header) — used by IDE/desktop clients like Cursor, VS Code, Claude Desktop, and CLIs:
```
Authorization: Bearer YOUR_API_TOKEN
```
Get your token from the Stealth API dashboard.

Tools

Tool	Purpose	Key arguments
`generate_content`	Start a full Stealth Agent run (research → draft → optional fact-check → humanize → optional images).	`preset` (`academic` \| `seo` \| `social`), `prompt`, `platform` (for `social`), `enableFactCheck`, `enableImageGeneration`
`humanize_text`	Start a humanization run for existing text.	`text`, `qualityMode` (`fast` \| `quality`), `model` (`heavy` \| `lite`), `outputFormat` (`text` \| `markdown`)
`get_run_status`	Get the current status and final result of a run.	`runId`

generate_content and humanize_text return a runId immediately (the work runs asynchronously). The agent then calls get_run_status with that runId until the status is completed, failed, or cancelled.

On a completed humanization, howLikelyToBeDetected is a 0–100 human-likeness score where higher is better (more human, less likely to be flagged as AI).

Install

One-click for Cursor:

The one-click button adds the server URL; add your Authorization: Bearer YOUR_API_TOKEN header in the generated mcp.json. (Cursor connects with a bearer token. For a no-token, one-click sign-in use claude.ai (web) with OAuth — see below.)

Or configure manually for your client:

Cursor
VS Code
Windsurf
Claude Desktop
Other clients

Add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global), then reload MCP servers:

{
  "mcpServers": {
    "stealthgpt": {
      "url": "https://stealthgpt.ai/api/mcp/mcp",
      "headers": { "Authorization": "Bearer YOUR_API_TOKEN" }
    }
  }
}

Add to .vscode/mcp.json:

{
  "servers": {
    "stealthgpt": {
      "type": "http",
      "url": "https://stealthgpt.ai/api/mcp/mcp",
      "headers": { "Authorization": "Bearer YOUR_API_TOKEN" }
    }
  }
}

Add to your Windsurf MCP config (~/.codeium/windsurf/mcp_config.json):

{
  "mcpServers": {
    "stealthgpt": {
      "serverUrl": "https://stealthgpt.ai/api/mcp/mcp",
      "headers": { "Authorization": "Bearer YOUR_API_TOKEN" }
    }
  }
}

Add it as a custom connector (Settings → Connectors → Add custom connector) using the URL https://stealthgpt.ai/api/mcp/mcp and sign in with OAuth — same flow as claude.ai (web).Alternatively, bridge over stdio with a bearer token using mcp-remote:

{
  "mcpServers": {
    "stealthgpt": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://stealthgpt.ai/api/mcp/mcp",
        "--header",
        "Authorization: Bearer YOUR_API_TOKEN"
      ]
    }
  }
}

Any client that supports a remote (Streamable HTTP) MCP server with custom headers:

URL: https://stealthgpt.ai/api/mcp/mcp
Header: Authorization: Bearer YOUR_API_TOKEN

Once connected, the three tools appear in the client and the agent calls them when relevant — for example:

Humanize this paragraph: …

Generate an SEO article about home solar energy.

Example flow

The agent calls humanize_text with your text → receives { "runId": "…", "status": "queued" }.
The agent polls get_run_status with that runId.
When status is completed, it returns the humanized result, howLikelyToBeDetected, and billing fields (wordsSpent, creditsSpent, remainingCredits, billingMode, meteredChargedCredits).

Connect in claude.ai (web)

The MCP server is available as a custom connector inside claude.ai — connect once with one-click OAuth (no API token to paste):

Open connector settings

In claude.ai, go to Settings → Connectors → Add custom connector.

Add the server

Set a name (e.g. StealthGPT) and the URL:

https://stealthgpt.ai/api/mcp/mcp

Click Connect, sign in with your StealthGPT account, and approve access on the consent screen. The connector switches to Connected.

Use the tools

Ask Claude to generate or humanize content, e.g. “Humanize this paragraph…” or “Generate an SEO article about home solar energy.” Claude calls the tools and polls the run for you.

OAuth and the bearer token resolve to the same StealthGPT account, so runs are billed identically regardless of how you connect.

ChatGPT

A ChatGPT connector / app uses the same MCP server and OAuth. It’s in progress and subject to ChatGPT’s connector requirements (workspace tier and tool conventions) — this page will be updated when it’s available.

Notes & limits

Runs are billed to the account that owns the API token, using the standard word-based billing.
Generation can take from seconds (humanize-only, lite) to a few minutes (full academic/seo pipeline). Because it’s async, the tool call returns quickly and the agent polls.
See the async workflow and Stealth Agent async runs pages for the underlying REST contract.

​Server endpoint

​Authentication

​Tools

​Install

​Example flow

​Connect in claude.ai (web)

​ChatGPT

​Notes & limits

Server endpoint

Authentication

Tools

Install

Example flow

Connect in claude.ai (web)

ChatGPT

Notes & limits