We've written a utility, mcp-shell, for integrating your shell with Claude using the Model Context Protocol. Check out the source code or just run npx mcp-shell.
A lot has been happening under the surface at High Dimensional Research.
Careful readers may have noticed slight changes to our website as we enter our next era, including some new faces in the shop. We'll have more to say as we go, but our same basic mission continues in a new form: whereas before we "connected and extended new intelligence," we now "build the infrastructure for intuition." What does this mean?
We've historically thought of models and intelligences as aspirations toward an extension of the human mind -- not necessarily literal, in-your-brain extensions (though who knows what the future will bring), but as extensions of yourself to handle and organize information. When we move our bodies, we simply "intuit" what to do. The command is not announced to ourselves, but inferred. Interface design, at its best, turns intention into an intuitive gestalt. Likewise, when we work with machines, at their best they extend our own intuitions into the world. Likewise, how else could we describe the probabilistic mechanisms of a generative model as anything other than "intuitive"? Models don't high-level reason so much as ... infer, slot in a little at a time. And so we've refined our mission toward the broader goal of enforcing that relationship with new intelligence, both its and our intuitive leverage.
mcp-shell
In the meantime, we've been experimenting with Anthropic's recent Model Context Protocol and built a basic utility to help Claude Desktop users integrate Claude with their own desktop: a package called mcp-shell.
Just run npx mcp-shell and Claude Desktop will automatically pick up the server and notice it can run a shell as a local function call.
Of course, we've put in some basic safeguards to avoid disastrous results, but it should enable you to work with your own files in a fun way.