MCP Server

The DocuPipe MCP server gives your coding agent ready-made DocuPipe tools. Tell the agent what you want from a document, and it uploads the file, extracts the fields, waits for the result and saves the JSON. Nobody has to write API calls.

MCP (Model Context Protocol) is the open standard coding agents use to plug in outside tools. It works with Claude Code, Codex, Cursor and other clients that can start a local MCP server.

Before you start

A DocuPipe account. Sign up if you don't have one.
Node.js 18 or newer on your computer. The server runs there through npx.
Your API key, from the Developers card on the Settings page.

👍
The Developers card has an MCP server section with a copy button for Claude Code, Codex and Cursor. It copies the install line with your API key already filled in.

Install it

Replace your-api-key with your key.

Claude Code

claude mcp add docupipe --scope user -e DOCUPIPE_API_KEY=your-api-key -- npx -y @docupipe/mcp

Codex

codex mcp add docupipe --env DOCUPIPE_API_KEY=your-api-key -- npx -y @docupipe/mcp

Cursor and other clients that use an mcp.json file

In Cursor, add this to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "docupipe": {
      "command": "npx",
      "args": ["-y", "@docupipe/mcp"],
      "env": { "DOCUPIPE_API_KEY": "your-api-key" }
    }
  }
}

🚧
Your API key gives full access to your account, and these install steps save it in the client's settings on your computer. If you put an mcp.json file inside a project, keep it out of version control.

Use it

Start a new session in your agent, open the folder that holds your documents, and ask in plain words:

Use DocuPipe to extract the invoice number, date, vendor and line items from every PDF in ./invoices and save one JSON file per invoice.

If you don't have a Schema yet, DocuPipe drafts one from the document. The result includes its schemaId, so the agent can reuse it and get the same fields from similar files. To control the fields yourself, name them in your request or ask the agent to use a schema you already saved.

More things you can ask for:

"Classify the PDFs in this folder using my DocuPipe classes."
"Split scans.pdf into one document per invoice."
"List my DocuPipe workflows and run the invoice one on this file."
"Convert statement.pdf to an Excel file and save it here."

Tools

Tool	What it does
`extract`	File or URL in, structured JSON out. Uploads, extracts with a saved schema, an inline JSON Schema or a schema DocuPipe drafts, waits, and can save the result
`upload_document`	Upload a local file or URL and wait for it to be read
`get_document`	Status, page count and parsed text of a document
`standardize`	Extract from a document that is already uploaded
`get_standardization`	Fetch an extraction result
`create_schema`, `list_schemas`	Save a JSON Schema, or reuse one the account already has
`list_classes`, `classify`	Classify a document into the account's classes
`split_document`	Split a file into sub-documents, with AI choosing where each one starts
`list_workflows`, `run_workflow`	Run a saved Workflow on a file
`convert_to_excel`, `get_excel_conversion`	Turn a document into an `.xlsx` workbook without a schema, and save it locally
`get_job`, `wait_for_jobs`	Check or wait on any job

Credits

The tools use credits from your account, the same way the API does.

What the agent runs	Credits
Upload of a PDF, image, Word or HTML file	1 per page
Upload of a CSV, Excel, JSON, text or XML file	Free
Extraction at standard effort	2 per page
Extraction at high effort	4 per page
Extraction with citations	3 per page
Drafting a schema when you don't provide one	1 per page, on at most the first 5 pages
Saving a JSON Schema	Free
Classification	0.1 per page, rounded up per document
Splitting	0.2 per page, rounded up
Conversion to Excel	2 per page

A workflow run costs the upload plus each step in the workflow. See Understanding Credits and Billing for every operation.

Long documents and large files

Each tool call returns within about 45 seconds. When DocuPipe needs longer, the tool returns status: "processing" with every ID created so far and the exact call to resume with. The agent makes that call, and resuming never uploads or extracts a second time.

Local files can be up to 100 MB. Pass a URL for anything larger.
Relative file paths start from the folder the server was started in, which for most clients is your project folder.
Results longer than about 40,000 characters are written to a docupipe-results/ folder there, and the tool returns the path with a preview.

Where your data goes

Your documents are uploaded to your DocuPipe account, exactly as they are through the API, and the same storage and retention settings apply. The results the tools return are passed to your coding agent, so they also fall under the terms of the agent and model you use.

📘
The server runs on your computer, so it needs a client that can start a local program. The web versions of claude.ai and ChatGPT can't use it, because they need a hosted server with a login, and DocuPipe doesn't have one.

Troubleshooting

What you see	What to do
`DocuPipe rejected the API key`	Check `DOCUPIPE_API_KEY` against the key on the Settings page
`Not enough credits`	Your account has run out of credits. See what happens when you run out
`npx` isn't found, or the server won't start	Install Node.js 18 or newer, then start a new session in your agent
`Not found` on a result	The extraction may still be running. Ask the agent to check its job with `get_job`

To use DocuPipe from an agent without installing anything, see Using DocuPipe from Claude Code, Codex or Cursor.