AI 어시스턴트를 위한 문서 처리 초능력

⚡ Endpoint

The MCP server runs over Streamable HTTP at:

https://api.documentize.app/mcp

The server uses stateful sessions. After the initial initialize request the server returns an Mcp-Session-Id header; include it in every subsequent request. MCP clients handle this automatically.

🔌 Connect from Claude Desktop

Add the server to claude_desktop_config.json (%APPDATA%\Claude\ on Windows, ~/Library/Application Support/Claude/ on macOS):

{
  "mcpServers": {
    "documentize": {
      "type": "http",
      "url": "https://api.documentize.app/mcp",
      "headers": {
        "Authorization": "Bearer dmk_your_key_here"
      }
    }
  }
}

Restart Claude Desktop. The Documentize tools will appear in the tool list. Omit the headers block to connect anonymously — the tools still work, but under the anonymous daily limit.

🔌 Connect from VS Code (GitHub Copilot)

Create .vscode/mcp.json in your workspace (or add to User Settings):

{
  "servers": {
    "documentize": {
      "type": "http",
      "url": "https://api.documentize.app/mcp",
      "headers": {
        "Authorization": "Bearer dmk_your_key_here"
      }
    }
  }
}

Open GitHub Copilot Chat, switch to Agent mode, and click the Tools button — Documentize tools will be listed there.

🔌 Connect from Cursor / Cline

In Cursor, open Settings → MCP and add a new server:

{
  "name": "documentize",
  "type": "http",
  "serverUrl": "https://api.documentize.app/mcp"
}

Cline users can add the same entry under MCP Servers in its settings panel.

⚙️ How tasks work

All processing tools are asynchronous. Each call starts a background job and returns a JSON object with a folderName field — that is your task ID.

1. Call a processing tool (e.g. ConvertDocument) → receive a task ID (folderName).
2. Call GetTaskStatus with that task ID and poll until statusCode is 200.
3. When complete, the response contains a sharedFiles array. Each entry has a fileName (display name) and an uploadFileName (a pre-signed S3 download URL). Open any uploadFileName URL in your browser to download the result file.

// 1. Start a task
ConvertDocument(fileUrl: "C:\\Users\\Alice\\Documents\\report.pdf",
                inputType: "pdf", outputType: "docx")
// → { "folderName": "abc123", "statusCode": 204 }

// 2. Poll until done (statusCode 204/202/203 = still processing)
GetTaskStatus(taskId: "abc123")
// → {
//     "statusCode": 200,
//     "sharedFiles": [
//       {
//         "fileName": "report.docx",
//         "uploadFileName": "https://s3.amazonaws.com/...presigned-url..."
//       }
//     ]
//   }

// 3. Download the result
// Open uploadFileName in a browser, or tell the agent:
//   "Download https://s3.amazonaws.com/...presigned-url... and save it as report.docx"

Use GetNextAppSuggestions after any operation to get recommended follow-up tools based on the output format.

📂 Providing your files

Most users are pointing the agent at a file on their own machine, so lead with that — a plain local path or a file:// URI works directly, no upload step required. The server also accepts remote URLs when that's where the file actually lives.

Local file path (most common)

File paths are resolved by the server, not by the client. They work whenever the MCP server can see the path you give it — Documentize SelfHost, a local development instance, or an agent running alongside the server. Just point it at the file:

// Windows absolute path
ConvertDocument(fileUrl: "C:\\Users\\Alice\\Documents\\report.pdf",
                inputType: "pdf", outputType: "docx")

// macOS / Linux absolute path
ConvertDocument(fileUrl: "/home/alice/documents/report.pdf",
                inputType: "pdf", outputType: "docx")

You can also use the standard file:// URI format:

// file:// URI — Windows
ConvertDocument(fileUrl: "file:///C:/Users/Alice/Documents/report.pdf",
                inputType: "pdf", outputType: "docx")

// file:// URI — macOS / Linux
ConvertDocument(fileUrl: "file:///home/alice/documents/report.pdf",
                inputType: "pdf", outputType: "docx")

Remote URL

Pass any publicly accessible https:// or http:// URL instead, and the server fetches the file automatically before processing.

ConvertDocument(fileUrl: "https://example.com/report.pdf",
                inputType: "pdf", outputType: "docx")

Supported file-source schemes

The server recognizes these URL schemes for fileUrl / fileUrls — anything else is rejected before a task is started:

private static readonly HashSet<string> SupportedSchemes = new(StringComparer.OrdinalIgnoreCase)
{
    "http", "https", "ftp", "ftps", "file", "local"
};

• file / local — a path on disk, either as a bare path (C:\..., /home/...) or a file:// URI.
• http / https — a remote URL the server downloads before processing.
• ftp / ftps — a file served from an FTP/FTPS host.

Practical agent prompts

Refer to files the way you naturally would — a path on your machine works directly:

// In Claude Desktop or Cursor chat:
"Convert my file C:\Users\Alice\Downloads\invoice.pdf to Word format"

"Compress file:///home/alice/documents/thesis.pdf and send me the download link"

"Extract all the text from /home/alice/documents/Q1.pdf"

If the file isn't local — it's already hosted somewhere — pass its URL instead: "Convert https://example.com/invoice.pdf to Word format".

👤 Accounts, limits & quotas

Usage is metered per account, and an MCP client is metered exactly like the website: one operation = one processing call, whatever the file count. Checking task status and asking about your own account are free.

Three tools let an agent manage this itself. None of them consume quota:

• GetAccountStatus — current tier, plan, operations used and remaining, reset time, and the safe parallelism for batch work.
• GetSignInInstructions — the exact steps to register, issue a key, and configure the client. The agent relays these to you; it cannot sign in on your behalf.
• SignOut — how to disconnect. Pass revokeKey: true to also revoke the key this connection uses (permanent, and it breaks every client sharing that key).

A healthy, signed-in connection looks like this:

GetAccountStatus()
// → {
//     "ok": true,
//     "account": {
//       "authenticated": true,
//       "authMethod": "mcp_key",
//       "credentialRejected": false,
//       "tier": "paid",
//       "plan": "Business",
//       "unlimited": true,
//       "maxParallelRequests": 8
//     },
//     "advice": "Unlimited plan. Mass processing is supported; use ProcessBatch
//                with up to 8 parallel requests."
//   }

A free account running low looks like this:

GetAccountStatus()
// → {
//     "account": {
//       "authenticated": true, "tier": "free", "unlimited": false,
//       "used": 4, "limit": 5, "remaining": 1,
//       "resetsAtUtc": "2026-08-28T00:00:00Z",
//       "maxParallelRequests": 2
//     },
//     "advice": "1 of 5 operations left today (resets 2026-08-28 00:00:00Z)..."
//   }

When a call is refused

Refusals come back as a structured object with a machine-readable error and a nextStep for the agent to follow, so it stops instead of retrying blindly. Each code means something different:

// Example refusal — note that it is explicitly not retryable
CompressDocument(fileUrl: "C:\\Users\\Alice\\Documents\\big.pdf")
// → {
//     "ok": false,
//     "error": "credential_rejected",
//     "statusCode": 401,
//     "message": "The Documentize API key in this MCP client's configuration was
//                 refused — it has been revoked, deleted, or was copied incorrectly...",
//     "nextStep": "Tell the user to sign in ... issue a new one, and replace the value
//                  in this MCP client's configuration (then restart the client).",
//     "retryable": false
//   }