Maxime Forest

Hi @Confused Hao

I've created a test app that replicates the basic features of the AI Assistant of RDM so you can test the fix I have running in my local RDM and validate that it works with a DeepSeek API key.

I'll send you a private message on the Forum, in that private message you will have a link to https://send.devolutions.net which will have an attachment (.zip file).

We can continue exchanging in private messages after this until we figure out the solution to get it working.

Hi @Confused Hao

I won't be able to test the full end-to-end integration with the DeepSeek API token since I don't have a subscription to the DeepSeek Platform, but rest assured I'll find a way to get this working for you.

I was able to test the DeepSeek model in another way using OpenRouter, and I applied fixes to get it working properly, but that alone doesn't guarantee it will work directly with a DeepSeek API token. Once I'm pretty confident in the fix, I'll let you know which version of RDM has it or provide you a custom RDM's installer for you to test.

I'll keep you informed when I have more details.

Best regards,

Thanks for the information!

Give me some time to investigate and see what I can do. It might take a bit of time since I don't have direct access to the DeepSeek Platform. I'll do tests on other similar platforms like OpenRouter, which I can configure in the AI Assistant using the same custom provider. OpenRouter has DeepSeek models, which I have access to.

I'll keep you informed as soon as I have more details.

Hi @Confused Hao,

I'm Maxime, AI Software Architect at Devolutions. I work on the MCP server and the AI Assistant integration for RDM.

I'll gladly help you with the issue you're facing. I created the custom provider for a specific use case for a client who was using a specific AI inference provider. Therefore, it's possible that there are issues with other products that I didn't test. Would you mind giving me more information about the product (website) you're using and how you've configured the custom providers (Anthropic, OpenAI with Response or Chat Completion API, Ollama), and which model you're trying to use?

Best regards,

As soon as we have the Devo CLI stable enough for preview, I’ll let you know so you can try it out, see if it fits your use case, and give feedback.

Best regards,

Hi @pico

I'm Maxime, AI Software Architect at Devolutions. I work on the MCP server and the AI Assistant integration for RDM, and soon, an unreleased tool.

Let's start by going over what RDM's MCP Server can and cannot do. We offer many tools through the MCP server. I've got a couple of tools that are important to the flow you're discussing in order to retrieve the Session entry in RDM: RDM_list_datasources, RDM_list_vaults, RDM_search_entries, RDM_get_entry. You can list the datasources (now called workspaces—just not renamed in the MCP tools yet), list your vaults, search for entries (list entries), and use the last one to get the full details of an entry.

We allow sending sensitive information through the MCP response (e.g., passwords) only if the user allows it in their RDM settings (AI Assistant > MCP Server > Exclude sensitive [...] unchecked). With that said, if you uncheck this, you must be aware that the AI model will have access to the sensitive information. Depending on the provider used, it might write it to a file. I've seen the Copilot provider do that in cases where the length of the data received is too large, in order to aggregate after. I don't think I've seen this behavior with Anthropic or OpenAI providers.

I feel like RDM MCP could be working for you if—and only if—the security aspect is something that you have leverage on. Unchecking the "Exclude sensitive [...]" checkbox means knowing that the data will include sensitive information for the AI. If I may suggest, letting any AI provider access passwords is far from ideal and is not something I would recommend.

You could have RDM open and use Devolutions Agent on the machines, and drive the MCP Server to connect to the machine and answer "What does the system log say since this morning? Any critical errors?" We have an SSH terminal entry that you could use for this use case, which can be driven by the MCP server. I've used it multiple times for demos where I show the top 5 processes running on a machine. The IP, hostname, and passwords never leave RDM using that method; all you need is the entry's ID, which you can find using the RDM_search_entry tool.

---

This brings me to the next part, an unreleased tool we have been developing that is not yet ready for customers: Devolutions CLI. The CLI we're working on will have a syntax similar to this:

devo secret run --env-ref HOST=devo://…/field/host --env-ref PASS=devo://…/field/password -- <tool>

This will allow an Agent to run the Devo CLI to obtain sensitive information (password) and store it in environment variables. We're building it for a use case with IronRDP, where we want to connect to a remote machine without passing the credentials in plain text, driven by an AI model.

---

Let me know if I was clear enough or if there's something else that needs clarification. I'm here to help.

Best regards,

Thanks @tobiasroeder for sharing your setup, it could help other users wanted it to try it also. If you happen to stumble upon bugs while using it don't hesitate to report them.

Best regards,

Hi,

I was not able to get it done in time for today's release. It will most likely go into the next one, for which I don't have a date yet. I'll keep you posted!

Hi!

RDM was released today, and it includes the addition of GitHub Enterprise URL! Please let me know if everything is working as expected on your end, and I'll close this topic.

RDM 2016.2.17.0

Release notes: https://devolutions.net/remote-desktop-manager/release-notes/

Hi!

RDM was released today, and it includes the fixes! Please let me know if everything is working as expected on your end, and I'll close this topic.

RDM 2016.2.17.0

Release notes: https://devolutions.net/remote-desktop-manager/release-notes/

Hi @tobiasroeder,

Quick update for you, since it's been 18 days since my last reply. I've got a working solution with Open WebUI running in Docker locally, as you can see from the following screenshots. I'll do my best to get it done and tested so it can go into the next release on Monday, August 3rd.

Best regards,

Quick update, if everything goes as planned, RDM should have a new update next Monday (August 3rd) that should fix your issue.

Best regards,

Quick update, if everything goes as planned, RDM should have a new update next Monday (August 3rd) with the support for GHE.

Best regards,

Hi @noeth,

Let me introduce myself first. I'm Maxime, an AI Software Architect. I am in charge of the MCP server and the AI Assistant in RDM.

I'm taking over this ticket from my colleague Carl. The feature for GHE has already been implemented internally and is ready to ship. If everything goes as planned, RDM should release a version next Monday (August 3rd).

I'll be closing this topic as it is a duplicate of a feature request from another RDM user: https://forum.devolutions.net/topics/52487/the-option-to-change-github-copilot-url-to-githhub-enterprise-url. Feel free to continue the conversation there!

Best regards,

Hi @dominikmueller,

I was able to do manual testing for 80% of what I needed without my own GHE setup. I'm confident that it should work for the remaining 20% that I haven't tested manually, since the flows are very similar to logging in with GitHub. I would rather have you try what I have done so far than make you wait longer for the infra setup to be completed.

I'm in the process of bringing this feature into the upcoming 2026.2.X release. I'll have this tested with our QA team and will let you know when I have information about the release date.

@danieliwaszkiewicz I've had the issue again since around 11 AM. I found a way to handle the schema / contract not being respected when the tool is called so that it just works. I'm about to get the fix reviewed, tested by QA, and ready for a future 2026.2.X version of RDM. I'll keep you posted when I have more information on the release date.

@danieliwaszkiewicz Windows Claude 1.22209.3 (babe11) 2026-07-19T02:06:35.000Z.

@danieliwaszkiewicz I decided to test it again this morning in Claude Desktop and at my surprise it is working on both 2026.2.14 and the internal future 2026.3.0 version. Can you test it again on your end after updating your Claude Desktop?

@danieliwaszkiewicz Sorry for the delay, I was testing various use cases all afternoon and looking at logs of Claude Desktop before I reply here.

I was able to reproduce the issue you're getting with Claude Desktop not being able to interact with opened connections in RDM. I tested it also with 2 other apps that can connect to MCP server (RDM with AI Assistant & VS Code) and both had no issues getting me the "top 5 processes" from an SSH Linux VM.

The issue is really with Claude Desktop only, even the Claude Code CLI can execute SSH commands using the RDM MCP tools.

Claude Desktop formats the JSON input sent to the MCP server tool differently which explains the error.

I'll have to do a workaround specifically for Claude Desktop and that specific MCP tool to make it work, I'll be working on this tomorrow. I'll try to let you know when I have an ETA of this fix.

@danieliwaszkiewicz Let me see if I can reproduce it with the steps provided. I'm currently installing 2026.2.14 and the exact jetsocat version locally, and will test it with Claude Desktop. I'll also check the other bugs you've mentioned (MCP server settings not getting saved).

I can answer this in the meantime:

a full Claude Desktop restart is required

When RDM closes or crashes, the child process which holds the MCP server is killed. Claude Desktop has no retry mechanism implemented on their end when a MCP server becomes unreachable; it ends up stopping the connection and staying like that. The only way to get it back is, like you said, to restart Claude Desktop. They have an open issue on their GitHub repository https://github.com/anthropics/claude-code/issues/54136 since April 27th.

Hi @danieliwaszkiewicz,

We've recently released a new version of RDM (2026.2.14) on July 15th, which includes the fixes for the MCP tools.

I was on vacation last week; otherwise, I would have given you an update earlier!

Let me know if everything works as expected on your end after you update, and don't hesitate to reach out.

Best regards,

Hi @dominikmueller,

I worked on this last week, I've got the flows completed in RDM but I still need to setup myself to do a full test manually with a GHE url in our environments.

I'm on vacation this week, I'll be working on this next week.

Best regards,

Hi @tobiasroeder,

I’ve been working on implementing the Custom Provider this week. I now have a functional AI Assistant running locally through OpenWebUI, using both GPT-5.6 from OpenAI and Ollama models.

I’m close to completing the feature for version 2026.3 (next RDM version), with plans to backport it to a 2026.2.x release. However, I’ll be on vacation next week, so it will take a little longer than expected to finalize and deliver the feature.

I'll keep you posted as soon as I get an idea on when this will be released!

Best regards,

Hi @tobiasroeder,

I'm Maxime, AI Software Architect at Devolutions.

I had a similar request for another software that proxied OpenAI endpoints. OpenWebUI appears to do something similar with Ollama models.

I think the existing code would support it. The only missing piece in the RDM and AI Assistant configuration seems to be the ability to pass the Authorization header, which is required for OpenWebUI’s Ollama API proxy support: https://docs.openwebui.com/reference/api-endpoints#-ollama-api-proxy-support. I don’t currently support this for any of the AI Assistant providers in RDM.

I have one or two ideas in mind. One of them is a custom provider that could have a custom name, header configuration, such as Authorization in your case, a URL or endpoint, and the provider it proxies, such as Ollama, Anthropic, OpenAI, etc. This way, I could support any product that proxies providers.

I’ll create a feature request on my end and keep you posted.

Best regards,

I'll do a full validation of every connections types that supports sending commands through the send_command MCP tools. I've investigated a bit this afternoon and found some issues that are easily fixable on my end that will most likely fix all of them at the same time.

Thanks for reporting this!