Claude Code's MCP Problem Just Got Fixed

Summary

In this video, the speaker discusses a recent update to Anthropic's Cloud Code and MCP servers, highlighting a new MCP search tool that improves performance by preventing the preloading of tool definitions. This change reduces context window bloat and enhances agent efficiency. The speaker also compares various tools and configurations, including Docker's MCP toolkit and the use of skills and sub-agents, to demonstrate how the lazy loading mechanism works. Ultimately, the update promises better scalability for Cloud Code, enabling more tools without negatively affecting performance or context space.

Takeaways

• 😀 The MCP search tool in Cloud Code solves the issue of context bloat caused by preloading all MCP tool definitions into the context window.
• 😀 Lazy loading of MCP tools ensures that tool definitions are only loaded when needed, optimizing the context window and improving performance.
• 😀 The MCP Launchpad, an open-source tool for managing MCPs, is compared to the new MCP search tool, with the latter being more efficient and integrated into Cloud Code.
• 😀 Preloaded MCP tools were consuming significant token space (22,000 tokens or more), affecting the performance of Cloud Code.
• 😀 Users were previously trying to use skills and subagents to hide MCP tool definitions, but this didn't work as MCP tools still impacted the main context window.
• 😀 Docker's MCP Toolkit also didn't fully solve the context bloat issue, as it still resulted in all tool definitions being loaded into the context window.
• 😀 Cloud Code now includes an 'enable tool search' setting in the `settings.json` file, which must be activated to enable lazy loading of MCP tools.
• 😀 Once the tool search setting is enabled, the context window is cleared of MCP tools, freeing up space and making the system more efficient.
• 😀 The MCP search tool works by selecting tools through a search query, either via direct tool name or using keywords to find related tools.
• 😀 As the conversation progresses, the list of loaded tools grows, but the lazy loading mechanism ensures only the tools needed at that moment are actually loaded into the context.
• 😀 Compacting the conversation will clear all loaded tools from the context, forcing Cloud Code to reload them when necessary, which helps to optimize the use of context space.

Q & A

What issue does the video address regarding MCP servers in cloud code?

The video highlights the problem of MCP servers being directly connected to cloud code, which degrades performance and consumes a large portion of the context window. This results in slower performance and inefficient use of available tokens.

What tool did the creator build to solve the context bloat issue, and how does it work?

The creator built the MCP Launchpad, which allows cloud code to search for tools without preloading tool definitions into the context window. This significantly reduces context bloat and improves performance.

What recent change did Anthropic make to address the MCP context bloat problem?

Anthropic released a new MCP search tool with Cloud Code. Instead of preloading all MCP tool definitions into the context window, Cloud Code now loads tools only when needed, improving efficiency.

What is the limitation of using skills or sub-agents to hide MCP tool definitions?

While skills and sub-agents can restrict which tools are accessible, they don’t prevent MCP tools from still being connected to the main cloud code context window, which means they still contribute to context bloat.

How does the MCP Docker toolkit work in relation to context bloat?

The MCP Docker toolkit simplifies connecting to Cloud Code by adding an MCP gateway, but it doesn’t solve context bloat. It still preloads all MCP tool definitions into the context window, which consumes tokens and space.

What steps must be taken to enable lazy loading of MCP tools in Cloud Code?

To enable lazy loading, users need to create or modify a settings.json file within the project’s .cloud folder and add the 'enable tool search' setting, ensuring that tools are loaded only when required, not preloaded.

What happens to the context window after enabling lazy loading of MCP tools?

After enabling lazy loading, the context window becomes free of preloaded MCP tools, and available space increases significantly, allowing for more efficient use of the context window during operations.

How does the MCP search tool work to load MCP tools dynamically?

The MCP search tool works by searching for and selecting specific tools as needed. It can either select tools directly by name or use keywords to find relevant tools. Once a tool is selected, its definition is injected into the system message, allowing it to be used in the conversation.

What impact does lazy loading of MCP tools have on the performance of long sessions?

Lazy loading improves performance in long sessions by preventing unnecessary tool definitions from accumulating in the context window. As tools are only loaded when needed, it reduces the context usage over time and maintains better performance.

How does compacting the conversation affect loaded MCP tools?

When a conversation is compacted, all previously loaded MCP tools are cleared from the context window, freeing up space. If needed, tools are loaded again as the conversation progresses, but the context window remains efficient and uncluttered.