- The MCP protocol allows AI to connect with external operating system tools and data in a standardized way.
- Server management is performed by editing the anythingllm_mcp_servers.json file and the Agent Skills interface.
- It is possible to control the automatic startup of servers to optimize the consumption of computer resources.
- Operation depends on the necessary binaries and commands being correctly installed in the host's PATH.
How to start and stop MCP servers from AnythingLLM Desktop? If you're looking to give your AI workflow an extra boost, you've probably heard of the Model Context Protocol, or MCP. This technology, powered by Anthropic, is essentially a bridge that allows language models to move beyond their own silos and interact with other languages. interact with external data and tools installed on your own equipment in a standardized way.
AnythingLLM Desktop has integrated this protocol so that its agents are not just experts at chatting, but can also perform real actions. From reading active PDF files in Bluebeam Revu to managing system processes, the ability to connect MCP servers It turns the app into a kind of command center for your operating system, allowing AI to know exactly what's happening in your other apps.
How to set up MCP servers
To start using these features, the first thing to know is that AnythingLLM doesn't start all servers when it opens, so as not to eat up all the RAM from your computer. By default, the system will automatically activate the MCP servers as soon as you navigate to the "Agent Skills" page or when you use the command @agent in the chat to ask for something specific.
If you want full control, you can manage power on and off from the interface. Just go to settings, enter agent skills, and, after selecting the server you're interested in, click on the gear icon to press the "Start" or "Stop" button. This is very useful if you have a server that consumes a lot of resources and you only need it occasionally.
JSON file configuration and management
The "magic" of MCP servers lies in a specific configuration file. If you look in the AnythingLLM storage directory, you'll find a folder called plugins where this file resides. anythingllm_mcp_servers.jsonThis is where you should add the objects from each server you want to integrate.
To add a new server, simply insert the server configuration following the MCP specification. If the server doesn't appear after making changes to the file, you don't need to restart the entire application; just click the button. Refresh in the agent's skills panel so that the system rereads the JSON and restarts the processes.
The Bluebeam Revu case study
A prime example of its power is the integration with Bluebeam Revu (requires Max plan and version 21.10+). For this to work, you must first enable the MCP in Revu's preferences by selecting AnythingLLM Desktop as the destination. Then, in the AI app, you replace the contents of the JSON file with the code provided by Bluebeam and update the server list.
Once connected, you can do amazing things like open a PDF in Revu and ask the agent: "How many pages does this document have?"The agent will consult the active file in real time. Don't forget to activate the options in the settings. generate files and summarize documents to take advantage of the full flow.
Technical details and optimization
It is essential to understand that AnythingLLM primarily supports transport. stdio (standard and simple) and also SSE or Streamable for real-time responses. A critical point is that the command you define in the JSON (such as npx, node, or uv) must be installed on your machine and available in the PATHbecause the application does not install external dependencies for you.
If you notice that a server starts on its own and slows down your PC, you can add the property anythingllm.autoStart: false in its JSON configuration file. This way, the server will remain asleep until you decide to wake it up manually from the user interface.
Problem solving and debugging
If you see that the server is not starting or the agent is not calling the tool, the first thing to look at is the status on the server card within the interface; if there's an error, it will appear there. If the problem persists, the best option is to check the desktop application's logs to see exactly what's failing.
Sometimes the problem isn't the server, but the language model. If you're using a very small local model With a limited context window, the AI may not be able to reason correctly when to call the MCP tool. In such cases, trying a more robust model usually solves the problem.
Tool maintenance and disposal
When you no longer need a server, you can delete it in two ways. The easiest way is to go to the interface, select the server, click the gear icon, and choose Remove MCP serverThis will delete the entry from the JSON file and kill the process immediately. You can also do this manually by editing the file and refreshing the UI.
Keep in mind that even if you delete the AnythingLLM server, the tools you installed using commands like uv tool install They remain on your hard drive. If you want to completely clean your machine, you'll need to manually delete those tools from the terminal.
Passionate about technology since he was little. I love being up to date in the sector and, above all, communicating it. That is why I have been dedicated to communication on technology and video game websites for many years. You can find me writing about Android, Windows, MacOS, iOS, Nintendo or any other related topic that comes to mind.