MCP Tools Just Got 10x Faster In Claude Code

The Claude Code team has addressed MCP's biggest issue by implementing tool search functionality that reduces context usage by up to 95%. Instead of preloading all available tools into the context window (which could consume tens of thousands of tokens), the system now searches for specific tool names before using them. This fundamental change solves the problem of token consumption occurring even before users write their first prompt. The video poses intriguing questions about why this wasn't the original implementation and whether the technique was inspired by Cloudflare's approach.

MCP servers have become ubiquitous across platforms including GitHub, Docker, Notion, and BetterStack, with widespread adoption in Claude Code and LLMs beyond just coding use cases. However, this popularity has exposed three critical problems: naming collisions, command injections, and most significantly, token inefficiency. The traditional approach of preloading all tools from connected servers into the model's context window creates visibility but at a massive cost. This preloading includes tool names, descriptions, and complete JSON schema documentation with parameters, types, and constraints, resulting in substantial data consumption before any actual work begins.

The Redis team's configuration demonstrates the severity of the token inefficiency problem with concrete numbers. Their setup utilized 167 tools from four different servers, consuming over 60,000 tokens before writing a single prompt. This represented almost half of Claude Opus' 200,000 token context window, and this calculation doesn't even include skills and plugins. For users with multiple servers installed, the token consumption could become even more substantial, severely limiting the available context for actual prompts and responses. This case study illustrates why token efficiency became the most pressing concern for MCP implementations.

While some models like Gemini offer larger context windows up to 1 million tokens, which might seem to solve the preloading problem, there's a critical performance trade-off. Models tend to perform worse as more content is added to their context window, regardless of the total available capacity. This means that simply having a larger context window doesn't eliminate the inefficiency problem. The degradation in model performance with increased context highlights why a more sophisticated approach to tool loading is necessary, setting up the explanation of potential solutions.

Two primary solutions have emerged in the community to address MCP token inefficiency: the programmatic approach championed by Cloudflare and the search approach implemented by the Claude Code team. The search process works with a threshold-based system: Claude first checks if preloaded MCP tools exceed 10% of the context window (20k tokens for a 200k window). If below this threshold, the system operates normally with all tools preloaded. When the threshold is exceeded, Claude dynamically discovers the correct tools using natural language processing and loads only 3-5 of the most relevant tools based on the specific prompt, allowing these selected tools to function normally while drastically reducing overall context consumption.

The tool search feature was the most requested enhancement on Claude Code's GitHub repository, and its implementation mirrors the progressive disclosure approach used in AgentSkills. Similar to how AgentSkills loads only skill names and descriptions initially, then fully loads specific skills when they're identified as relevant or mentioned in the prompt, the MCP tool search maintains minimal context overhead. Only when the system identifies tools as relevant does it load their complete specifications into the context window. Both Anthropic and Cursor have reported significant benefits from implementing this search-based approach, validating the technique's effectiveness across different platforms.

The programmatic approach represents a fundamentally different solution where models orchestrate tools through code execution rather than API calls. Instead of making individual API tool calls for sequential operations that depend on previous responses, Claude can write a Python script that handles all the orchestration logic. The script executes and presents results back to the model in a single operation. Cloudflare has advanced this concept further by having models write TypeScript definitions for all available tools, then executing the code in sandboxed environments (typically workers). This approach transforms the tool usage paradigm from declarative API calls to imperative code execution.

Interestingly, the Claude Code team tested the programmatic approach but ultimately found the search method to perform better, a surprising conclusion given Claude's well-known coding capabilities. The speaker expresses skepticism about this decision, noting that Vercel's agent browser CLI with headless Chromium works exceptionally well in Claude Code. The speaker suggests that converting all MCP tools into CLI commands using something like MCPorter could be more efficient and context-friendly, as models could run specific CLI commands for tools instead of loading tool specifications into context. Despite these alternative viewpoints, the speaker acknowledges the positive development that MCP server issues are being actively addressed, suggesting it might encourage broader adoption beyond single-server configurations.