Hi,
for anyone who wants to use BoltAI2 with the new DEVONthink MCP server, here’s the config that worked for me:
{
“mcpServers”: {
“devonthink”: {
“command”: “/opt/homebrew/bin/npx”,
“args”: [“mcp-remote”, “http://localhost:8420”, “–transport”, “http”]
}
}
}
Note, that you have to install node e.g. via “brew install node”.
Best,
Timo
Thanks for sharing this configuration. Though it’s specific to BoltAI, it could also help direct others using similar MCP-capable apps.
I am curious: What’s your use case for DEVONthink MCP and how is the experience working with it in BoltAI?
Right now I am just exploring the possibilities of the DT MCP server. For me it’s like the API I always wished DT had. I want to integrate my database with my Python tools and access information through Jupyter notebooks.
The BoltAI integration is working fine. No crashes or error during my 3h session. If you like I can send you the chat protocol between me, DT and Claude. It had some interesting remarks about the inner workings of DT.
One thing happened, though: Because I wasn’t paying attention, I spent 90 USD on Claude restructuring my tags etc. But this was clearly my own stupidity and nothing technical. Just as a warning 
90 dollars? 
How many items and tags do your databases contain?
Not that much, actually. See screenshot.
My mistake was to use Claude Opus with Max Effort. Total overkill.
For those of us waiting on the sidelines before dipping a toe in the AI waters, we salute your
a) willingness to experiment and
b) willingness to document your mistakes and misadventures here
so that we may learn from them and know what to avoid when we finally get into this MCP business.
I can’t speak for others but I would have to make some time to experiment with all this fabulous new stuff because I can’t risk a mission critical database on something where I am flying blind.
Omg. I’m glad I have turned off auto usage limits. It was turned off , but the limit was set to $500 
I have hit the usage quota limit a few times while reorganizing my DEVONthink tags and groups with MCP. Would have cost me a few hundreds probably… I’m on a $ 20 pro plan.
This config also works with llama-cpp server…
Start a local proxy with uvx and point it to Tomo’s config…
uvx mcp-proxy --named-server-config mcp-config.json --allow-origin "*" --port 8001 --stateless
You can install uvx with homebrew if you don’t already have it.
Then, start llama-cpp server with the --webui-mcp-proxy flag.
My launch line currently looks like this…
nohup llama-server \
--models-dir ./models/ \
--jinja -c 0 \
--host 127.0.0.1 --port 8033 \
--offline \
--models-max 1 \
--webui-mcp-proxy \
> "logs/llama-server.$(date +%Y%m%d_%H%M%S).log" 2>&1 &
Then in the llama-cpp server web ui, add the mcp server as…
http://127.0.0.1:8001/servers/devonthink/mcp
Now your local chats with local models should be able to connect to your local proxy and talk to your local DEVONthink MCP server. Local.
Side note: you’ll see I’ve got concurrent models loaded set to 1 because I’m substantially poorer than Sam and Dario and can’t afford more than 32Gb of unified right now. I comfortably switch between gpt-oss-20b-mxfp4 for general purpose (MoE model, actually really good given its size) and then qwen3.5-9b-mxfp4 for tight-context coding (small blocks / functions, not entire workspaces / repos).
I can actually run both of those at the same time if I up --models-max to 2. But my mbp starts to sound like an aircraft carrier and there’s nothing left in the memory bank for anything else.
I’ve tried working those models on larger contexts and with agentic workflows using pi. I don’t think they’re smart enough (although I think you can probably do a lot more than most realise by working on the harness, adding skills and mcp hooks).
Ideally though, you’d want one of the qwen 30b models or larger. (I can actually get that to run at a squeeze… but it’s far too slow for any practical use.)
To each his own… but there is also an immense opportunity cost to missing out on a transformational increase in the usefulness of your database.
Maybe I’m doing something wrong or forgetting something, but in my case, I get an error when I try to save the script.
Can you paste the config here as text? I’ll check it.
Below the plain text:
{
“mcpServers”: {
“DEVONthink”: {
“command”: “/opt/homebrew/bin/npx”,
“args”: [“mcp-remote”, “http://localhost:8420”, “–transport”, “http”]
}
}
}
I must say that I have found the right solution by now. I also connected the DEVONthink MCP to Cloud Desktop, and then I imported the MCP server in Bold AI 2 by selecting the three dots in he upper right corner. This gave me the following configuration, which does work:
{
“mcpServers” : {
“DEVONthink” : {
“args” : [
“–stdio”
],
“command” : “/Applications/DEVONthink.app/Contents/Library/LoginItems/DEVONthink MCP.app/Contents/MacOS/DEVONthink MCP”
}
}
}
With ChatGPT replacing the Codex app with a new ChatGPT app, the DEVONthink MCP menu was unable to auto detect the ChatGPT app as the Codex App, and so the checkbox for installing to Codex was greyed out.
But I was able to use the method described in this thread to manually install the MCP into the new ChatGPT app.
It might be nice to make these instructions plain in the DEVONthink MCP menu? Bookends does the same. They have the buttons to auto-install it in the recognized apps (Claude, ChatGPT, LM Studio, etc.) but then also have right next to it - here is the JSON you should use if you want to install manually in another app.
The next release will support this and also add buttons & menu items to copy the required config.
