File tools

Overview

Three built-in integrations cover file work, and they do different jobs:

• file-tools — read data out of files. Extract rows from CSV, Excel, and JSON, text and tables from PDFs, and content from Word documents, then export results back to Excel.
• file-generation — produce documents. Render PDF, DOCX, XLSX, PPTX, Markdown, JSON, and TXT from structured content blocks or a Jinja2 template.
• filesystem — read and write files on the server itself, restricted to paths an administrator allows.

Enable only the ones a project needs. Each is an MCP server the platform ships and manages, so agents discover the tools and call them during reasoning like any other tool.

file-tools

Extract structured data from files and export results to Excel. It exposes ten tools.

Getting a file in

• upload_file — upload a file from a local path and get back a file_id UUID. Pass that file_id to the extraction tools.
• preview_file — peek at a file's structure without extracting everything. Returns column names and a few sample rows, so an agent can decide what to do before pulling the whole file.

Extracting rows

• extract_rows — auto-detects the file type (Excel, CSV, or JSON) and returns rows as an array of objects. Use this when you don't want to care about format.
• extract_csv_rows — extract rows from a CSV file.
• extract_excel_rows — extract rows from an Excel file (.xlsx, .xls).
• list_sheets — list the sheets available in an Excel workbook.

PDFs and Word

• extract_pdf_text — extract text page by page. Returns an array of pages, each with page_num and text.
• extract_pdf_tables — scan every page for tables and return them as rows.
• extract_docx — extract content from a Word .docx document. Defaults to mode='tables' for table extraction.

Sending results back out

• export_to_excel — merge ForEach results, one row per input plus its output, and export as an Excel file uploaded to storage.

file-generation

Generate documents from structured content blocks. It exposes five tools.

• render_document — render a document and return its storage metadata: key, content_type, filename, and size. Supports PDF, DOCX, XLSX, MD, JSON, and TXT.
• render_from_template — render a Jinja2 template with supplied data and produce a PDF, HTML, or PPTX file.
• describe_document — read back a presentation you already generated, slide by slide, in the same shape an edit would take.
• revise_document — change a presentation you already generated without re-sending it. Pass the key returned by render_document.
• upload_branding_image — upload a header or footer image. Returns a key you can pass as header_image or footer_image on later renders.

Templates and branding

An administrator can pre-load DOCX, PPTX, and XLSX templates plus header and footer branding images when connecting the integration. Documents then render against your own house style rather than a default.

filesystem

Read and write files on the server, with path-based access control. Every file operation is restricted to directories an administrator has allowed.

Configuration

Use cases

Invoice data into a spreadsheet

A ForEach node walks a folder of invoices. For each one, extract_pdf_tables pulls the line items, an agent normalises them, and export_to_excel merges every result into a single reconciliation sheet.

Branded report generation

An agent gathers findings, then render_document produces a PDF with your header and footer images applied. describe_document and revise_document let a reviewer request changes without regenerating the file.

Contract review

extract_docx pulls the clause tables out of a Word contract, an agent checks them against policy, and a Human Task node routes anything uncertain to a person.

Next steps