Had DT and LMS working just fine…then I updated both apps to their latest versions and updated my Mac Studio to 26.5.1. Now DT doesn’t “see” LMS and LMS logs show unexpected endpoint errors from DT. Seems that LMS is expecting “POST /api/v1/chat” but receiving “GET /api/v1/chat”. Using port 8420 for all of this. Also, I think LMS update from REST V0 API to V1 which might play into this. Any feedback appreciated. Very new to LMS and AI so just searching for a little firm ground to push forward.
The support for LM Studio hasn’t been changed recently in DEVONthink. We will check this but a screenshot of Settings > AI > Chat might be useful.
Just updated to LM Studio 0.14.5 build 2 but everything is still working as expected. Therefore screenshots of your settings would be appreciated.
I had that same issue with latest LMS versions. Communication cuts suddenly and no log in DT and no log in LMS as well. It simply stops. So and then, even with 64K context enabled at both sides, answers cut in, say 20 lines of text without any error. It seems like the socket pipe loses the information. I don’t discard a bug in macOS socket framework, as so and then I get empty Safari pages when loading from a local server in same Mac.
Which model do you use? Do you experience similar issues with Ollama?
I don’t have Ollama installed.
Muy most used model is QWEN3.6 35B 8bit uncensored heretic, that I use for almost all. I don’t know the exact name because I’m at work now and that is from home. It runs in my new M5 Max 128GB of RAM, the memory is not an issue. Anyway, other models have the same issue. I will be more specific this evening.
I use some scripts in Python with same model via LMS, mostly for book translation and style improvement with very heavy usage (hours) and same 64K window without any disconnection.
Christian, those are the ones I use, by order of usage: qwen3.6-35b-a3b-uncensored-heretic-mlx, qwen/qwen3.5-35b-a3b, google/gemma-4-31b
And I just did a summary of a web captured PDF and nothing at DT side after some time of progress bar. No message in log. I’m attaching the LMS log of the operation. And prepare an email with the things that DT generate when requesting support.
Below are a few images of my LMS and DT setups, and what happens upon DT starting.
LMS is up and happy, so I open DT.
As soon as DT starts the “unexpected endpoint…” errors are logged. I get the same results with Apple Intelligence enabled or disabled, the only difference being with AI enabled I can click on the Chat window and enter keyboard input. With AI disabled I cannot enter any input.
I would appreciate your thoughts regarding this. Thank you for the assistance.
I have replicated the issue with 9 different models (both GGUF and MLX). I have not tried this with Ollama. In the screenshots I just posted I was using Mistral 7B.
What does LM Studio’s mcp.json file look like?
- That is lacking all the information to connect to DEVONthink’s MCP.
- And your picture of LM Studio’s Server Settings is wrong. That’s for LM Studio.
PS: How much RAM do you have on that Mac?
OK, I was certain something was wrong. Just did not know what. So I will take your information, downgrade to a previous version, and see if I can learn the details of what needs to be done. Thank you for your time and assistance. Oh, by the way, the Mac studio is 36 GB. Have a good evening.
You’re welcome and that’s a good amount of RAM to potentially have a useful experience.
PS: I posted this on another thread…
Do you actually use LM Studio in DEVONthink’s chat assistant or try to use DEVONthink (via MCP) in LM Studio’s chat? In addition, is the Launch at login option in Settings > AI > MCP enabled? Port 8420 is actually used by the MCP server by default.
I’ve got the same problem as OP since updating DT to 4.3.1 on my Mac mini. It always worked flawlessly in conjunction with LM Studio. (Use case: Models loaded in LM Studio in DT’s chat assistant)
Now, however, the only thing changed is updating from DT 4.3. to 4.3.1 and suddenly, the model picker is greyed out and no model can be picked for LM Studio (as in OPs screenshot).
What’s more, this must be a DT-related issue, since I have not yet updated my Macbook from 4.3. to 4.3.1 and the model picker works as it did before (my LM Studio server is accessible inside my LAN from all Apple devices).
Happy to provide any info needed to asses this!
Actually DEVONthink’s support for LM Studio in version 4.3 and 4.3.1 is identical. But you could try version 4.3 nonetheless:
https://download.devontechnologies.com/download/devonthink/4.3/DEVONthink.dmg.zip
Does downgrading fix the issue?
Yes, see my updated post above re my not yet updated MBP.
I still wonder whether downgrading the affected machine would make a difference. Maybe it’s related to the configuration or the macOS version or third-party software?
Actually, the MBP is still on Tahoe 26.5 (including DT 4.3) whereas the Mac mini is already on Tahoe 26.5.1 (running DT 4.3.1). The machines are, otherwise, configured mostly identical. I can try and Downgrade to DT 4.3. on the Mac mini. Would that be helpful?