An Experimental MCP Server for Tinderbox
Over the last several weeks, I’ve been making more and more use of Claude Cowork to aid in my professional and personal pursuits—more on that at the end. However, it seems that Tinderbox’s native MCP functionality isn’t compatible with Cowork’s use of containers for executing code and connecting to MCP servers. So, out of necessity, I built an enhanced MCP server for Tinderbox. This MCP server not only provides the same tools as the native one, it also bundles reference material about adornments, expressions, action code functions, built-in attributes, and export codes.
While Tinderbox’s native MCP server is good, it doesn’t provide any guidance to the AI on how it ought to interact with the richness of Tinderbox. The lack of guidance was a design decision by Tinderbox’s author, Mark Bernstein, to use as little of the AI’s context as possible. Fortunately, we were able to solve this by creating a succinct collection of resources instructing the AI on how to use Tinderbox. These resources were derived from Mark Anderson’s excellent aTbRef.
By default, the enhanced—but experimental—MCP server provides some basic Tinderbox guidance to the AI. The guidance is approximately 3,400 tokens. So, if our context window is 200,000 tokens, the initial instructions consume about 1.7% of it—a very modest amount. On top of the initial guidance, we make six discrete resources available to the AI:
| Resource | Tokens | Description |
|---|---|---|
| action-functions.md | ~8,250 | 300+ action functions by category |
| export-codes.md | ~6,340 | 46 export template codes |
| expressions.md | ~3,840 | Expression and action code syntax |
| system-containers.md | ~3,210 | Prototypes, Templates, Hints, Composites |
| action-attributes.md | ~2,770 | 12 action-holding attributes |
| adornments.md | ~1,650 | Map adornments and smart adornments |
| Total all resources | ~26,060 |
The AI is instructed to load these resources as needed. But even in the worst case, with all resources loaded, the server stays under 15% of a 200K context window, leaving over 170,000 tokens for the conversation.1
So what have I been able to do with this new Tinderbox MCP server and Cowork? Well, just this week, I had it help me with:
- Contact management. Created people notes for those I interact with regularly at work using the company’s directory. I was able to enrich the data by having Claude pull additional information about each individual based on their LinkedIn profile.
- Reference wrangling. In support of a research project that I’m pursuing, I had several people send me articles that I ought to read. I had Claude parse these emails and create notes in Tinderbox with the complete reference information and abstract.
- Competitor analysis. I set up a job in Claude Cowork to scour the web regularly, performing competitor research. These findings are saved and updated in Tinderbox, one note for each competitor.
Without the AI’s help, it would have taken me many more hours to complete these tasks on my own. But even better, Claude and the MCP server performed well. Not once did I need to correct it or consult external Tinderbox documentation.
-
I imagine context window management is going to become largely a non-issue for nearly everyone soon, as the frontier AI labs, including Anthropic, continue to roll out 1 million context windows in their consumer products. ↩︎