Update Claude Code with New Tool Search Feature

Anthropic updates its Claude Code system with a new feature called Tool Search, allowing artificial intelligence (AI) agents to access tool definitions dynamically. This feature is particularly useful for optimizing the use of the Model Context Protocol (MCP), an open-source standard released by the company in 2024 that connects AI models to various tools and data sources.

The new feature was announced last night, representing a significant change in the architecture of Claude Code. With Tool Search, agents no longer need to consult lengthy documents before carrying out a specific task, which consumes a high number of tokens, a limited resource defining the amount of information that can be processed.

Prior to the update, these agents often “read” long manuals for each tool, which could waste up to 33% of the context window before a prompt was even sent. By introducing lazy loading functionality, Anthropic allows agents to search for tool definitions only when necessary.

Initialization Rate for Agents

The previous implementation of the MCP allowed connections to over 50 tools, but as the ecosystem grew, what became known as the "initialization rate" emerged — an implicit cost faced by developers. Thariq Shihipar, a member of the technical team at Anthropic, noted that complex setups could consume 67,000 tokens, dramatically limiting users' operational capacity.

As reported by Aakash Gupta, author of an AI newsletter, this setup forced developers to balance between limiting their tools to a few options or seeing a large part of the context budget compromised even before starting their tasks.

How Tool Search Works

With the new feature, Claude Code monitors context usage in real-time. If the tool documentation threatens to exceed 10% of the total available, the system changes strategy and loads a lightweight search index. When requested, the system searches the index and retrieves only the relevant definition.

Internal testing by Anthropic demonstrated a drastic reduction in token usage, dropping from 134,000 to just 5,000 when using the new feature. Shihipar emphasized that server instructions are now a crucial part of MCP, allowing Claude to know when and how to search for tools.

Lazy Loading and Increased Precision

Besides the savings in tokens, one of the most important benefits of this update is the ability to focus. When the AI's context is saturated with irrelevant information about tools, its performance in reasoning and argumentation can be impaired. Boris Cherny, director of Claude Code, noted that the new feature provides more focus and improves response accuracy.

Benchmarking indicates that the accuracy of the Opus 4 model in MCP evaluations increased from 49% to 74%, while the accuracy of Opus 4.5 jumped from 79.5% to 88.1%. The focus on relevant information helps the model concentrate its attention mechanisms on user inquiries.

Maturity of Infrastructure

The update represents an evolution in handling AI infrastructure. The primary engineering challenge shifts as systems scale, and regulated practices in software development fields are now essential. Aakash Gupta discusses the evolution of integrated development environments, noting that modern tools do not load extensive definitions initially, and this should be applied to AI agents.

Implications for the Ecosystem

For the end user, this update provides a smoother experience, but for the developer ecosystem, it opens up new possibilities. With Tool Search, there is no longer a rigid limit on an agent's operational capacity. Thousands of tools can be included without penalties as long as they are effectively utilized.

This change transforms the context economy, from a scarcity model to an access model, potentially revising what it means to have tool-rich agents. The update is already being rolled out to all users of Claude Code, and developers are encouraged to adopt the ToolSearchTool to support this new dynamic loading model.