No response when using the OpenAI api provided by LM Studio

[pks888]

Description

After I entered the URL provided by lmstudio in DBeaver's AI settings and clicked 'Test Connection', it just keeps sending request. But actually, if I try navigate to 'http://127.0.0.1:1234/v1/models' directly in Edge, Edge can display the model list results. I tried changing the port number and turning off the firewall, but it still doesn't work.

DBeaver Version

25.2.4

Operating System

win11

Database and driver

No response

Steps to reproduce

No response

Additional context

No response

[ShadelessFox]

Hello @632575987,

Try replacing 127.0.0.1 with localhost to see if it has any effect.

[pks888]

Hello @632575987,

Try replacing 127.0.0.1 with localhost to see if it has any effect.

I've already tried using localhost, but it doesn't work either.

[ShadelessFox]

There may be something useful in the logs. Can you please provide them? You can find instructions here: https://github.com/dbeaver/dbeaver/wiki/Log-files

[darrylsepeda]

Hi @ShadelessFox

I've just encountered same issue, but no logs appear on that part/section:

Basically, the process just stuck there, LM Studio is running fine and callable from the browser, but not from Dbeaver.

[ShadelessFox]

@darrylsepeda don't use the error log view. Instead, collect and send us the log files mentioned in the linked article.

[darrylsepeda]

Hi @ShadelessFox ,

Apologize for slow response.
is this the one dbeaver-debug.log:

you could ignore the empty API token at the beginning, InterruptedException because its already stuck on loading window for a long time so I ended up canceling the loading window.

[Bader-CN]

Hello All

I have the same issue. This is my test, hoping it could be helpful for troubleshooting.
I found that vLLM can be connect, but LM Studio cannot connect it.
And there’s no record of the requested entry in LM Studio’s logs.

here is my logs:
2025-11-30 15:33:32.594 - InterruptedException
java.lang.InterruptedException
at org.jkiss.dbeaver.model.ai.utils.MonitoredHttpClient.send(MonitoredHttpClient.java:74)
at org.jkiss.dbeaver.model.ai.engine.openai.OpenAIClient.getModels(OpenAIClient.java:114)
at org.jkiss.dbeaver.model.ai.engine.openai.OpenAIEngine.getModels(OpenAIEngine.java:57)
at org.jkiss.dbeaver.ui.ai.preferences.AIPreferencePageMain.lambda$2(AIPreferencePageMain.java:266)
at org.jkiss.dbeaver.ui.UIUtils.lambda$17(UIUtils.java:1942)
at org.eclipse.jface.operation.ModalContext$ModalContextThread.run(ModalContext.java:123)
2025-11-30 15:33:37.284 - Error reading model list
org.jkiss.dbeaver.DBException: Request was cancelled
at org.jkiss.dbeaver.model.ai.utils.MonitoredHttpClient.send(MonitoredHttpClient.java:91)
at org.jkiss.dbeaver.model.ai.engine.openai.OpenAIClient.getModels(OpenAIClient.java:114)
at org.jkiss.dbeaver.model.ai.engine.openai.OpenAIEngine.getModels(OpenAIEngine.java:57)
at org.jkiss.dbeaver.ui.ai.engine.openai.OpenAiConfigurator.fetchOpenAiModels(OpenAiConfigurator.java:201)
at org.jkiss.dbeaver.ui.ai.model.CachedValue.get(CachedValue.java:36)
at org.jkiss.dbeaver.ui.ai.engine.openai.OpenAiConfigurator.lambda$3(OpenAiConfigurator.java:157)
at org.jkiss.dbeaver.ui.ai.model.ModelSelectorField.refreshModelList(ModelSelectorField.java:116)
at org.jkiss.dbeaver.ui.ai.model.ModelSelectorField$2.run(ModelSelectorField.java:105)
at org.jkiss.dbeaver.model.runtime.AbstractJob.run(AbstractJob.java:119)
at org.eclipse.core.internal.jobs.Worker.run(Worker.java:63)
Caused by: java.lang.InterruptedException
at org.jkiss.dbeaver.model.ai.utils.MonitoredHttpClient.send(MonitoredHttpClient.java:74)
... 9 more

I guess it might be that DBeaver’s request method is incompatible with LM Studio.
causing LM Studio to reject the request, which is why DBeaver gets stuck in the request state.

[ShadelessFox]

Can you turn on this option to see if it has any effect? It's available since 25.2.5.

[Bader-CN]

@ShadelessFox

Is same issue when selected "Use legacy API"

[adrenalin8231]

LM Studio does not handle HTTP/2 requests well. lmstudio-ai/lmstudio-bug-tracker#1079
There is not much DBeaver can do: preferring HTTP/2 over 1.1 is actually good behavior.
If you are not on the metered network, you can try ngrok. Or try reverse-proxy: I managed to do it with Caddy

:11234 {
    # Accept HTTP/1.1 & HTTP/2 (including h2c)
    protocols h1 h2 h2c
    reverse_proxy localhost:1234 {
        # MITM: Strip upgrade headers so backend doesn't hang
        header_up -Upgrade
        header_up -Connection
        
        # Force backend connection to be standard HTTP/1.1
        transport http {
            versions 1.1
        }
    }
}

[HavenCTO]

Have you tried jan? https://github.com/janhq/jan/ it seems to have the same functionality as LM Studio but with HTTP/2 support

[jurandifr]

Hi everyone,

I was running into this exact same issue and wrote a lightweight Python reverse proxy to act as a bridge between DBeaver and LM Studio.

The root of the problem seems to be how DBeaver's strict HTTP client handles the streaming responses and headers from LM Studio. The proxy fixes this by intercepting the traffic, forcing "stream": false in the JSON body, stripping problematic chunked/encoding headers, and explicitly injecting the exact Content-Length so DBeaver knows exactly when the response ends (which prevents the infinite connection hang). It also prevents the ECONNRESET spam on LM Studio's side.

If anyone needs a quick workaround while this isn't officially fixed, I published the script here:
https://github.com/jurandifr/router-Lm-Studio

Just run it locally and point DBeaver's API URL to the proxy's port (e.g., http://127.0.0.1:8080/v1). Hope it helps someone!

[kevan]

大家好，

我遇到了完全相同的问题，并编写了一个轻量级的 Python 反向代理，作为 DBeaver 和 LM Studio 之间的桥梁。

问题的根源似乎在于 DBeaver 严格的 HTTP 客户端如何处理来自 LM Studio 的流式响应和标头。代理通过拦截流量来解决这个问题，强制 JSON 正文中的“stream”属性设置为 false，移除有问题的分块/编码标头，并显式注入精确的 Content-Length 值，以便 DBeaver 能够准确知道响应何时结束（从而防止无限连接挂起）。它还能防止 LM Studio 端出现 ECONNRESET 垃圾邮件。

如果这个问题尚未正式修复，需要快速的临时解决方案，我已将脚本发布在此处： https://github.com/jurandifr/router-Lm-Studio

只需在本地运行，并将 DBeaver 的 API URL 指向代理的端口（例如，http ://127.0.0.1: 8080/v1）。希望对大家有所帮助！

thanks ! so much.

[E1izabeth]

Is it still reproducible in the latest DBeaver version?

[kezarjg]

This is still reproducible, and I think I found the root cause. It is the HTTP/2 theory from @adrenalin8231 above, but it is fixable on the DBeaver side.

AbstractHttpAIClient creates its client with HttpClient.newHttpClient(), which defaults to HTTP_2. Over plaintext http:// there is no TLS and therefore no ALPN, so the JDK falls back to an h2c upgrade and attaches these headers to every request:

Connection: Upgrade, HTTP2-Settings
Upgrade: h2c

LM Studio, and uvicorn/httptools-based servers such as vLLM, treat that as an unsupported upgrade and stop parsing the request. The application then sees an empty body: vLLM answers POST /v1/responses with

1 validation error:
{'type': 'missing', 'loc': ('body',), 'msg': 'Field required', 'input': None}

I could only test vLLM directly, but the same upgrade headers would explain the LM Studio hang and the InterruptedException in MonitoredHttpClient.send reported in this thread. HTTPS endpoints negotiate over ALPN and never send those headers, which is exactly why only self-hosted plaintext endpoints are hit.

Pinning the shared AI client to HTTP/1.1 fixes it without a reverse proxy. HTTP/2 buys nothing for these APIs over a local unencrypted connection, so there is no real trade-off. I have a patch that does this and have verified it against vLLM: the body now arrives and the request completes.

Could this be reopened? I would like to submit the fix as a PR against devel.