What You Can Build When Your AI Agent Has 51 File Tools

You are mid-session in Cursor. The landing page is coming together, but the hero image is 4800 pixels wide and 3 MB. You need it at 1200px, 800px, and 400px, compressed, in WebP. Three options. You could open a browser tab, upload the image three times to some online tool, and download three results. You could write a quick shell script with ImageMagick. Or you could tell the agent: “resize this to three widths, compress to 80% quality, convert to WebP.” No tab switch, no script, no upload.

That third option is what the VaultTools MCP server makes possible. One package, 51 file processing tools, all running locally on your machine. This post walks through real scenarios where those tools save time during actual development sessions.

Resizing, Converting, and Compressing Assets Mid-Session

Landing pages, README screenshots, social preview cards. Every project eventually needs image work. With the MCP server connected, your agent can resize images to exact dimensions, compress them to a target quality, and convert between formats. No browser tab. No separate workflow.

The typical flow: you describe what you need (“hero image at 1200px wide, 80% quality, WebP”), and the agent calls the tools in sequence. The resized, compressed, converted file appears in your project directory. You stay in the editor.

This is especially useful for responsive image sets. Instead of manually generating three sizes, you describe the breakpoints and let the agent handle the repetition.

Automating Contract and Report Assembly

Imagine you have ten PDF chapters for a technical report. They need to be merged in order, with page numbers added, then compressed for email. Doing this manually means finding a PDF tool online, uploading every chapter (to someone else’s server), downloading the result, then repeating for page numbers and compression.

With the MCP server, the agent calls merge, then page numbers, then compress. Three operations, one prompt, zero uploads.

Every file stays on your machine. The MCP server uses file paths, not file bytes. Your PDFs, contracts, and reports never enter the AI context window and never leave your disk.

This matters most when the documents are sensitive: legal contracts, financial reports, medical records. The privacy guarantee is structural, not a policy promise.

Stripping Metadata Before You Share or Commit

Screenshots and photos carry more data than you think. EXIF metadata can include GPS coordinates, camera serial numbers, timestamps, and operating system details. Committing test fixture images with embedded location data is a real privacy leak.

The EXIF viewer and stripper lets an agent inspect metadata first, then strip it. In practice, you can ask the agent to check all images in a directory for GPS data and clean any that have it. No manual inspection of each file.

This is particularly relevant for open source projects where contributors submit screenshots or sample images. A quick metadata scan before merging a PR takes seconds with the MCP server.

Converting Between Formats Without a Script

Data format conversions are the kind of task that feels too small to write a script for, yet too tedious to do by hand. The MCP server covers the common ones.

Need to turn a CSV export into JSON fixtures for a test suite? CSV to JSON handles it. Comparing a YAML config with its TOML equivalent? YAML/JSON/TOML converter makes the round trip. Want to decode a JWT from a failing request without pasting it into jwt.io? The JWT decoder runs locally.

You can also chain these with other tools. Ask the agent to convert two config files to JSON, then run a text diff to spot the differences. That entire sequence stays inside your editor session.

Generating Build Artifacts Agents Would Otherwise Skip

AI agents are good at writing code but tend to skip asset generation. They will happily scaffold a web project without creating a favicon, generating QR codes for staging URLs, or computing checksums for release binaries.

With the MCP server, these become one-line requests. Generate a favicon set from the project logo. Rasterize an SVG to PNG for email clients that do not render vectors. Create a QR code pointing to the staging deployment. Calculate SHA-256 checksums for every binary in the release folder.

These are small tasks individually, but together they represent the gap between “the code works” and “the project is ready to ship.”

Reading and Transforming PDFs Without Parsing Them Yourself

PDF manipulation comes up more than you would expect during development. Documentation builds, invoice processing, test data preparation.

The MCP server gives agents the ability to split a document by page range, inspect metadata like author and creation date, add watermarks to outgoing files, and convert pages to images for visual regression tests or thumbnails.

All of this runs through the same Rust processing engine that powers the browser tools. No Ghostscript dependency, no Python subprocess, no external API.

One Line to Add It to Any Agent

Installation is a single command:

npx @vaulttools/mcp init

The interactive setup detects your AI tools and lets you pick a directory. For non-interactive use:

npx @vaulttools/mcp init --tool claude-code --dir ~/Documents

The --allowed-dir flag sets a security boundary. The server can only read and write files within the directories you specify. Nothing outside that boundary is accessible.

The MCP server works with any client that supports the Model Context Protocol: Claude Desktop, Claude Code, Cursor, Windsurf, and others. Configuration is a one-time JSON snippet in your client’s settings. See the full setup guide for details.

51 tools. One install. Every file stays on your machine.

What This Changes

The real cost of file processing during a development session is not the processing itself. It is the context switch. Opening a browser, finding a tool, uploading, downloading, renaming, moving files back into the project. Each interruption is small but they compound.

When your agent has direct access to these tools, file processing becomes part of the conversation. No tab switching, no uploads, no privacy trade-offs. You describe what you need, and the files appear in your project directory.

Try it: install the MCP server, connect it to your editor, and ask your agent to do something with a file. You will not want to go back to the browser tab workflow.