Local vs Remote MCP Servers — Transcript
Full transcript
- 0:00Pretty much all MCB servers you might
- 0:02have come across are local MCB servers.
- 0:05These are programs usually written in
- 0:07NodeJS or Python that run on your
- 0:10machine and let you enhance area apps
- 0:12like GitHub copilot cursor or cloud
- 0:16desktop. In some cases, like integration
- 0:18with your file system or a database
- 0:21running locally on your machine, it's
- 0:23necessary that the MCB server runs
- 0:26locally cuz it's impossible for a
- 0:28program running on the internet to be
- 0:30aware of your local files. But in other
- 0:32cases where you need access to external
- 0:35APIs like an MCB server for Sentry,
- 0:38which is an application monitoring tool,
- 0:41it gets a bit cumbersome when the MCB
- 0:43server runs locally. So when you want to
- 0:45ask C-Pilot about the back in your app
- 0:48to report to Sentry, Copilot first talks
- 0:50to the Sentry MCB server and locally
- 0:53which then in turns makes an API request
- 0:56to the Sentry API. There's a couple
- 0:58problems with this approach though. You
- 1:00have to create and manage API keys by
- 1:02yourself as opposed to using the
- 1:04standard all flows. It is a backend MCP
- 1:07server itself. You need to update the
- 1:09local MCP instance to get the bug fix.
- 1:12And if you got many MCP servers running
- 1:14locally, it can potentially be a
- 1:17performance bottleneck. However, the MCP
- 1:20standard also defines remote MCB
- 1:22servers. And this is not something new.
- 1:24It's actually been part of the MCB spec
- 1:26since day one. Although it recently got
- 1:29some pretty significant changes, but
- 1:31more on that in a bit. As you might have
- 1:33guessed, remote MCB servers don't run on
- 1:35your machine, but they run in the cloud.
- 1:38One good example of a remote MCP server
- 1:40is Git MCP. The name confused me a bit
- 1:44at first as it's not about git per se
- 1:46but about providing context about a
- 1:48GitHub repository of your choice to the
- 1:51LLM. So let's say we want to add context
- 1:53about the playright GitHub repo to
- 1:55C-Pilot. So we can ask questions about
- 1:58playright. We can grab Rimmel MCB for
- 2:00that add it to VS code and when we run a
- 2:03query we get back a result. Although I
- 2:06admit that for this example, we would
- 2:08have gotten the correct reply back even
- 2:10without any MCP servers. But the
- 2:12important bit here is that co-pilot
- 2:15talked directly to the MCP server
- 2:16running in the cloud with no additional
- 2:19servers running locally. If you want a
- 2:21remote server with authorization, it
- 2:24gets a bit more tricky for now as
- 2:26support for that is kind of lacking in
- 2:28most MCB clients, at least at the time
- 2:31of recording this video. One example is
- 2:33Sentry. And yes, they do have an
- 2:36official remote MCP server, which is far
- 2:39better than the local Sentry MCP server
- 2:41I've shown earlier. If you try to
- 2:43install it in the same way as a
- 2:45Playright Kit MCP server from before,
- 2:48you'll get a 401 error when you try to
- 2:50start it. Ideally, VS Code would launch
- 2:53a Sentry authentication page so you
- 2:55could authorize the Sentry MCB server,
- 2:58but this hasn't been implemented yet.
- 3:00again, at least at the time of recording
- 3:02this video. But there's an open feature
- 3:04request in VS Code's GitHub repo for
- 3:07that. What you can do, however, is to
- 3:09use an MCP remote MPM script, which is
- 3:13sort of a proxy for remote MCB servers
- 3:16that can handle all as well. This means
- 3:19that from Copilot's perspective, we're
- 3:22talking to a local MCP, but the local
- 3:24MCP is only a thin proxy layer to a
- 3:27remote MCP, which you'll be able to
- 3:29remove once C-Pilot has support for
- 3:32authenticated remote MCPS. And yes, I
- 3:35admit it's a bit convoluted for now, but
- 3:38it's the early days for the model
- 3:40context protocol, and support for some
- 3:42features is lacking in many clients.
- 3:45Anyways, for now to use the remote
- 3:47sentry MCP server, we'll need to install
- 3:50it by prepending the npx mcp remote
- 3:54command before the remote mcp server
- 3:56URL. And then when the server starts, a
- 3:59sentry off page launches, which is much
- 4:02better and safer experience than
- 4:04generating and storing the API keys. So
- 4:08yeah, as you saw, the remote MCB servers
- 4:11can offer better UX for off, better
- 4:13security, and less load on your machine.
- 4:16But the problem is that although they've
- 4:19been part of the MCP spec since day one,
- 4:21the client support for them is kind of
- 4:23lacking and there are only a handful of
- 4:26remote MCB servers available. And also
- 4:28at the end of March 2025, the remote MCP
- 4:32protocol has been changed from requiring
- 4:34a stateful HTTP connection and using
- 4:37server send events to a stateless
- 4:40protocol with simple get and post HTTP
- 4:43requests and optional server send
- 4:45events. And let me know if you're
- 4:47curious about the details of the new
- 4:48protocol. I might make a separate video
- 4:51on that. But for today, that's all I've
- 4:53got. Take
About this transcript
This page contains the full transcript of Local vs Remote MCP Servers by Tom Dohnal, generated from the public captions YouTube serves with the video. The transcript has 829 words across 118 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.