Remote MCP / AI detection
Ask an MCP-compatible assistant to analyze supplied text without leaving the conversation. The remote server returns structured detector details through one focused tool.
export DETECTING_AI_API_KEY="YOUR_API_KEY"
[mcp_servers.detecting_ai]
url = "https://api.detecting-ai.com/api/mcp"
bearer_token_env_var = "DETECTING_AI_API_KEY"
tool_timeout_sec = 300One shared remote MCP server
Detector and humanizer use the same Streamable HTTP endpoint. This page focuses on the detect_ai_text tool.
https://api.detecting-ai.com/api/mcpMCP tool contract
detect_ai_textAnalyzes user-supplied text for patterns associated with AI writing. The optional version input defaults to Detector v3, and the output includes sentence-level details plus the processed word count.
Default: version = "v3"
| Input | Type | Required | Description |
|---|---|---|---|
| idempotency_key | string | Yes | A 16–128 character URL-safe operation key. Generate one per intentional analysis and reuse it only for an exact retry. |
| text | string | Yes | Text to analyze. The MCP tool requires at least 10 characters. |
| version | string | No | v1, v2, or v3. The default is v3. |
{
"success": true,
"details": {
"chunks": [
{ "text": "...", "type": "AI", "score": 0.87 }
],
"ai_percentage": 42.5
},
"version": "v3",
"words_processed": 120
}Authentication
Pick the authentication path your MCP client supports. Never add credentials to the server URL itself.
Add only the MCP endpoint in a hosted connector. The client discovers the authorization service, opens sign-in and consent, then stores its own access token.
Store the API key in an environment variable or secret store. The client sends it as an Authorization bearer token when it calls the remote server.
Inside the conversation
The tool call stays in the active assistant session, which makes it easier to discuss the result beside the text that produced it.
Ask the assistant to check a draft, passage, or submission and explain that the result will support a human review.
The client sends the text to the remote MCP server with v3 by default and uses the detector balance on your API subscription.
Read the overall signal and sentence chunks, then discuss the result alongside drafts, citations, and other evidence.
MCP clients
Hosted Claude and ChatGPT connectors use OAuth. Codex, Claude Code, and similar local clients can use a private API key as a bearer token.
Add the remote endpoint as a custom connector and complete the browser authorization flow.
Use a custom MCP connector with OAuth when the required connector settings are available.
Reference the server URL and an environment variable that contains the bearer token.
Add the Streamable HTTP server with an Authorization bearer header in the local client.
Where it helps
MCP reduces context switching. The assistant can call the detector, receive structured data, and continue the review in the same working session.
Check a contributed draft, then ask for a passage-by-passage reading of the returned detector details.
Use the signal as one part of a conversation about sources, revision history, and the institution's policy.
Add a detector step to a controlled content workflow and route uncertain cases to a person for inspection.
MCP FAQ
The server is remote, usage comes from your API subscription, and each tool returns structured output to the client.
It is a remote Streamable HTTP MCP server that gives compatible assistants access to the detect_ai_text tool. The same server also exposes a separate humanize_text tool.
Hosted Claude and ChatGPT connectors can use OAuth. Codex, Claude Code, and other local-config MCP clients can connect with a Detecting AI API key sent as a bearer token.
It returns detector details, the version used, and the processed word count. The public Detector v3 compatibility contract uses raw 0–1 chunk confidence; ai_percentage is the share of input characters covered by AI-labeled sentences, not an average of chunk scores.
No. The result is a probabilistic signal. It should guide a closer review that also considers drafts, sources, writing history, and context.
Successful calls deduct the input word count from the detector balance on the same subscription used by the REST API.
One server, two focused tools
Read the complete setup guide, compare API plans, or inspect the other tool on the same remote server.