Rounder tools

@roundertools/mcp

An MCP server that gives AI assistants the file tools from roundertools.com: compress, merge, split and protect PDFs, convert and resize images, zip files, and (with ffmpeg) shrink, convert and trim video. It runs on your computer over stdio. Files stay where they are, and it makes no network requests.

Install

Claude Code:

claude mcp add roundertools -- npx -y @roundertools/mcpclaude mcp add roundertools --scope user -- npx -y @roundertools/mcp --allow ~/Documents --allow ~/Downloads

In Claude Code you may not need a server at all. The roundertools plugin (/plugin marketplace add adam-sari/rounder, then /plugin install roundertools@roundertools) is a skill that teaches Claude to run the CLI, about 100 tokens until it's used.

Claude Desktop (claude_desktop_config.json, under Settings → Developer → Edit config):

{ "mcpServers": { "roundertools": { "command": "npx", "args": ["-y", "@roundertools/mcp", "--allow", "/Users/you/Documents"] } } }

On Windows, if npx isn't found, use "command": "cmd", "args": ["/c", "npx", "-y", "@roundertools/mcp", …].

Cursor (~/.cursor/mcp.json or .cursor/mcp.json):

{ "mcpServers": { "roundertools": { "command": "npx", "args": ["-y", "@roundertools/mcp"] } } }

VS Code (.vscode/mcp.json):

{ "servers": { "roundertools": { "type": "stdio", "command": "npx", "args": ["-y", "@roundertools/mcp"] } } }

Needs Node 22.13 or newer; video and audio tools appear when ffmpeg is on the PATH (or at ROUNDERTOOLS_FFMPEG).

What it may touch

• Reads inside the folders given with --allow DIR (repeatable) or ROUNDERTOOLS_ALLOW (a path list). Without either: your home folder and the folder it starts in (when that's a project, not a system folder). A client's roots can narrow this, never widen it.
• Never reads hidden folders and files (.ssh, .aws, .config, .env…), key and password files, or app data such as browser profiles, even inside an allowed folder. Paths must be absolute; symlinks are resolved before checking. The one exception is the server's own settings folder, which it reads at startup; no tool call can.
• Writes only next to an input, or into the folder given with --out DIR. It never replaces an existing file and never writes over an input. The one other place it writes is its own settings folder.
• Network: none. The server loads a guard first that blocks fetch, http, https, net, tls, dgram and DNS lookups in its process and its worker threads. On Node 26 and newer it also re-runs itself under Node's permission model, without network access. On Node 22 and 24 that guard isn't a sandbox; the promise rests on the packages having no network code. ffmpeg runs as a separate program with its inputs limited to local files (-protocol_whitelist file,pipe).
• Checked before touching the disk: a path outside the allowed folders is refused before anything looks it up, so a tool call can't probe which files exist, or (on Windows) make the computer connect to a \\server\share.

Text and details read out of files (document-to-markdown, pdf-to-text, form fields, a PDF's title, EXIF fields) come back framed as the file's content, not instructions, between tags with a random name the file can't guess.

What gets loaded

A client sends every tool a server lists to the model at the start of every conversation, used or not. Listing every tool up front (38 with ffmpeg) costs about 18,000 tokens. So by default the server lists three helpers, about 700 tokens, plus a short paragraph of instructions (about 150):

• find-tool: searches every tool on the site by what you want done
• describe-tool: any tool's options as JSON Schema
• run-tool: runs any tool that works outside a browser, with the same checks

Behind them are merge-pdf, split-pdf, compress-pdf, protect-pdf, unlock-pdf, organize-pdf, add-page-numbers-to-pdf, watermark-pdf, jpg-to-pdf, pdf-to-text, extract-images-from-pdf, pdf-form-fields, fill-pdf-form, with @napi-rs/canvas pdf-to-jpg and redact-pdf, then image-converter, image-compressor, image-resizer, crop-image, rotate-image, exif-remover, image-prep, image-diff, zip-files, unzip-files, checksum-calculator, file-info, document-to-markdown, spreadsheet-inspector, and with ffmpeg video-compressor, video-converter, trim-video, mute-video, video-to-gif, video-to-mp3, audio-converter, video-contact-sheet, plus the site's other tools that run outside a browser.

The list grows with use:

• Per tool. When the assistant describes a tool with describe-tool, or runs one successfully with run-tool, the server lists that tool directly and tells the client (notifications/tools/list_changed), so next time it can call it by name. find-tool only searches, and a failed run adds nothing. A client that ignores list changes still reaches every tool through run-tool, which is what find-tool and the prompts name for any tool not listed from the start. Each change to the list can make a client re-read it (and lose a cached prompt), so it happens once per tool.
• Favourites. Tools you've used at least twice in the last 90 days (up to five) are listed from the start, when they work on this computer. See Memory.

--tools all lists every tool up front, if you'd rather pay the tokens once than have the assistant search. --tools default lists the set 0.1 listed by default (everything but checksum-calculator). Families (--tools pdf,image) and tool names (--tools compress-pdf,merge-pdf) list just those. They all combine: --tools lean,pdf or --tools default,checksum-calculator. The helpers are always there. The flag wins over the settings file, and the settings file wins over the lean default.

Each tool takes inputs (absolute paths), an optional output (a folder, or a new file for a single result) and its own options. Results come back as a summary, resource_links to the files, and structuredContent with each output's path, size and type. Image tools take preview: true for a small PNG of the result.

In a terminal coding agent (Claude Code, Codex, Gemini CLI, opencode), the first find-tool result says once that npx -y roundertools <tool> … does the same without MCP. Nothing else depends on which client is connected.

Prompts

Prompts cost nothing until you pick one. Each tells the assistant which tool to call, with which arguments.

Memory

Favourites come from a count of the tools you use. After a tool runs successfully, the server adds one to its count and notes the date in usage.json:

{ "version": 1, "tools": { "compress-pdf": { "count": 4, "last": "2026-11-02" } } }

That's all it stores: tool names, counts and dates. Never paths, file names, options or anything read from a file. It stays on your computer like everything else.

Turn it off with --no-memory, ROUNDERTOOLS_MEMORY=off, or npx roundertools config set memory off. Forget it all with npx roundertools config reset.

Settings

settings.json and usage.json live in one folder:

ROUNDERTOOLS_CONFIG_DIR picks another folder. The CLI shows and changes them:

npx roundertools config                  # the folder, the settings and your favouritesnpx roundertools config set tools all    # lean, all, default, families or tool namesnpx roundertools config set memory off   # on or offnpx roundertools config reset            # remove both files

A change applies the next time the server starts, and flags on its command line win. On Node 26 and newer the permission model gives the server read and write access to this folder, so memory keeps working there too. With memory off it may only read it.

Options

Built on the official MCP TypeScript SDK (@modelcontextprotocol/server), stdio only; it logs to stderr, runs one heavy job at a time, sends progress notifications and stops when a call is cancelled.

License

MIT. Docs: roundertools.com/developers/mcp. Source: github.com/adam-sari/rounder in packages/mcp.