I apologize for asking this since most of you seem to know what you are doing, but can anyone tell me what I need to do in Tinderbox to use Claude with it? I can’t find anything in TB 11.6 to hook it up. I’ve searched the forums but haven’t found anything that I recognize for accomplishing this.
I’ve been watching the meetup video with Andreas Grimm and it looks like the potential is amazing. I have a subscription with Perplexity and it is awesome, but I don’t know whether it will ever work with TB so I thought I’d try Claude. I do have AI integration enabled.
The forum is not the correct place to look for actual documentation. Had you first checked Tinderbox app Help or aTbRef, both cover this. Please read this aTbRef article and the sub-articles (you can ignore the one on Gemini). Short version:
Install Claude Desktop and then close.
In Tinderbox, use the Tinderbox 11 menu and select Enable All AI Integration.
Close Tinderbox.
Re-open Claude Desktop. It should re-open Tinderbox.
In any TBX wish to use in tandem with Claude, go to Documents Settings::General and enable AI use.
I took the plunge today and subscribed to Claude. Following the list of steps offered by @mwra , I was able to begin teaching myself how to use Tinderbox with Claude.
There were a few hiccups. One was the MCP server being offline. I quit both Tinderbox and Claude, relaunched Claude and it launched Tinderbox and things went ok from there.
I created a test document Claude Training, and began chatting with Claude.
I got tripped up in Hints/AI as Claude was working in the Claude container, while I was observing the Claude Code container.
There was also a genuine bit of mystery, as I distinctly recall prompting Claude set the OnAdd action of the Session Notes container to assign the Session Note prototype. Later, Claude asked me if I wanted it to do that. It’s a little unsettling.
I love that you’ve shared the session text, as seeing others edging forwards makes the initial plunge a bit less frightening if we erroneously the assume out first will have no missteps. It useful too to catch the tenor of the Ai’s replay and its degree of detail.
I’ve also had trouble with the MCP connection with Claude. I had it working and everything seemed fine, but then the connection dropped. I also quit both Claude and Tinderbox, relaunched Claude, it opened Tinderbox and everything seemed fine for a bit then the connection dropped again. Here is what Claude said:
Even get_document is failing now — the MCP connection dropped again entirely. This is clearly an instability issue worth solving properly.
I’d suggest posting in the Tinderbox Forum’s AI category (AI - Tinderbox Forum) describing the problem — that it connects initially but drops when switching documents or after a short time. Mark Anderson and the Eastgate folks are active there and this is exactly the kind of MCP/Claude integration issue they’d know how to fix. It may be something simple like a timeout setting or a configuration tweak.
I realize this isn’t the forum Claude suggested, but since the same issue came up here I thought it might be OK to post.
I’ve seen this, but it’s highly inconsistent. It also differs between Claude, Claude Code, and the Code tab in Claude Desktop.
I am reasonably sure it’s a Claude bug. Recently, Claude Code started to (presumably intentionally) shut down its Tinderbox and then reopen it later; I assume this is an effort to conserve computational resources.
If simply restarting isn’t helping:
Quit Claude and Tinderbox
Open Tinderbox
Toggle AI Integration off, then on again
Open the document of interest
In Document Settings, toggle AI Integration
Quit Tinderbox
Claude Code is able to see and diagnose its own settings, which can be helpful.
In recalcitrant cases, Claude’s Developer Preferences offers access to log files. I can take a look, or you can ask Claude to analyze them…
Don’t overlook the fact that Claude had some system outages recently. Unless you are working fully local, you are communicating with cloud-based systems and aspects of the internal at all sorts of levels fail or go slowly more often than we assume. I definitely experienced some Claude weirdness recently that appears to coincide with Claude system failures. There is a Claude page reporting performance (sorry, can’t find the link ATM, Google should know) and be aware vendors may <cough> not be the most honest reporters of their performance as they have stock-holders to keep happy. So, third-party reporters are probably more trustworthy sources.
Therefore, before rushing to re-boot everything, it pays to check Claude or [AI vendor] is operating correctly else all downstream test will fail to fix anything.
Dave, I agree with Mark Anderson, reading the transcript of your session with Claude spurred me to download it and try out the desktop app before I got into Tinderbox.
I wanted to see what it could do on its own. I have some data I collected from a scientific instrument that I exported from their software as a CSV file. The layout is really designed to work best with a single test, but I have batches of tests, each run with different materials created using a factorial type recipe, then tested using the same test at different temperatures. When it exports the sets of tests, it separates the columns of data with subheaders containing the test metadata (including the temperature). The way I create the x-y graphs that visualize the data is to stack the tests in a spreadsheet, and replicate the metadata for that test (ie the temperature) in their own columns, with rows matching those for the sample. Then I can use a Pivot Table to inspect the data, make sure I got all the samples I needed to test, or I can analyse it with Python and generate plots or statistics as I need.
Cleaning the data by copying and pasting takes me a long time; I paid my daughter the last time I did it and it took her a tenth of the time it would have taken me, but it still took hours. I wanted a Python script to do it, but the last time I did a similar script for a different instrument, it took me a couple months part-time to write the code. (Granted, I was also learning the programming language, and the other export had some very odd quirks.)
I described what I wanted to Claude, fed it the file, and in less than 10 minutes, it had written me a Pyhon script and output the modified file. The modified file is what I had asked for (though I plan to tweak it to remove extraneous information).
Speaking as someone who learned to program on punch cards, I have to say my mind was blown, so I am looking forward to trying Claude with Tinderbox, as I can see a number of possibilities for speeding up some boring aspects of reshaping/analysing both data files, and scientific literature.
Well done!
As an aside, if cleaning/reviewing data and not wanting to drop immediately to Python or R coding, take a look at the EasyDataTransform app from another small indie developer (I got a copy at the indie Summerfest sale).
It’s Java based so can run on Mac or windows and sits between spreadsheet and pure coding. I have trust issues with Excel (tends to re-format/damage data without warning) and going straight to code is a pain unless you are a seasoned coder.
This isn’t suggested as better than AI but may be useful for those who are interested in the quality of the data itself as opposed to looking for the fastest way to use the data for something else. i’ll admit Ii do more of the former and that usage is a bit niche.
Thanks, Mark, nice to know about this tool. I’ll bookmark it for future one-off uses, and I think my colleagues who don’t code may find it useful.
My work with instrument data is the very first step in Python scripts for statistical analysis and data plots. I only use Excel to make sure I have all the files in each set of conditions (a factorial design, if anyone cares about that), or to generate a quick statistical summary for someone else. I also share these scripts with other instrument users who do limited programming.
Using AI to do the job wasn’t my goal, it was a test case to challenge it with a task that I found tedious, and to see how generalized the script it emitted would be (Claude says it can be generalized by changing the column heads, but I have yet to test that with a different data file, so I’m sceptical). My next test will be to give it a data file from a different instrument that locked up scripts. It took me days to find a non-printing ASCII character (0xC-form feed) that wasn’t handled properly by my code. If it IDs that, I will be impressed.
If you know there are bad/unwanted/unexpected characters or encodings it is likely worth sharing that explicitly in the prompt. Even if you don’t know the exact nature of the badness you can give Claude some ‘serving suggestions’ of the type of thing to look for, as you might if briefing a human.
OK. I’ve been slow getting back to this but I have four college classes running concurrently and that is pushing me for time. I tried many things with Claude again and it gave me text to post here so you could see what is happening and what we have tried. Here it is:
I’m running Tinderbox 11 with Claude Desktop and have MCP integration enabled. I’m having a persistent issue where the connection appears to initialize correctly but breaks down when I try to do anything beyond get_document.
What works:
get_document (sometimes — returns the list of open documents)
set_document (returns OK)
What consistently fails:
get_notes (with any query)
get_view
All other tools
Setup:
Tinderbox 11 (latest version)
Claude Desktop (latest version)
macOS
MCP server configured in claude_desktop_config.json pointing directly to the Tinderbox 11 app binary
What I’ve tried:
Disabling and re-enabling AI integration in Tinderbox
Quitting and relaunching both Claude Desktop and Tinderbox
Letting Claude Desktop launch Tinderbox (rather than opening Tinderbox manually)
Closing all other Tinderbox documents so only the target document is open
Adding a timeout value (60000ms) to the MCP server config
Reverting to the original config
Pattern observed:
The connection seems to work briefly right after launch — get_document occasionally succeeds in that window — but then drops. The problem document (PandA Website Ideas 2025) has multiple maps, which may or may not be relevant.
Interestingly, when Claude Desktop launches, a macOS modal appears asking to install Command Line Developer Tools (for the git command). Not sure if this is related.
Any suggestions on what might be causing get_notes and get_view to fail, and how to stabilize the connection? I’d love to get this working seamlessly.
Double check that AI Integration is enabled both for Tinderbox as a whole (Tinderbox ▸ Enable Tinderbox Integration) and in your Tinderbox document (Document Settings: General). The best way to confirm this is to turn both settings off, then on.
Make sure neither Tinderbox nor Claude are running.
Start Claude Desktop. Use the Claude tab, not the CoWork tab.
Claude should launch Tinderbox. Be sure the desired document is open.
Ask Claude to read its Tinderbox Readings. This should teach Claude just enough about Tinderbox to get by.
If you get error messages, you’ll find a small expansion triangle in the right margin in Claude. Expand that to view the whole message; copy it and email it to bernstein@eastgate.com.
I use this pretty much nonstop, every day. The connection only drops when I quit one of the programs.