dsh-wolfram-bundle

If Wolfram Desktop or Engine is already on your machine and you want your AI assistant to evaluate Wolfram code, run tests, or read and write Notebooks directly, dsh-wolfram-bundle is the thin integration layer that makes it happen. It does not rewrite the MCP protocol and does not modify the official Paclet. Instead it fills three gaps: auto-discovering the local kernel, owning the process-tree lifecycle, and routing every side-effectful tool call through DSH's native approval chain. Once loaded, the model sees seven tools prefixed with mcp__wolfram__—arbitrary evaluation, test execution, Markdown-to-Notebook conversion, static inspection, symbol lookup, and more.

Security is the defining constraint. WolframLanguageEvaluator has no sandbox: it can write files, spawn processes, and hit the network. So it defaults to ask and will not fire until a human approves. Unknown tools shipped in future Paclet upgrades follow the same rule—ask first, not later. The supervisor terminates only the process tree it spawned itself; it will never reach for your running Mathematica front-end or servers owned by other clients. Even in the TerminateProcess window on Windows, orphan reclamation requires triple identity verification (image name, command line, creation-time window), and if any check fails the process is left untouched.

The target user is concrete: you run DSH as your daily AI coding assistant, have local Wolfram with AgentTools Paclet 2.2.0 or later, and want the model to safely evaluate a snippet or inspect a Notebook without silently writing files or spawning processes. Requires Node 22 or later; works on Linux, macOS, Windows, and WSL. Clean install, clean uninstall, no global dependencies.