Claude Code’s Biggest Upgrade Yet Fixes a Growing Agent Problem
The Claude code from Anthropic has just gotten one of the most wanted updates ever and in the process resolves a problem that was potentially suffocating the whole model ecosystem.
At the heart of this integration is the Model Context Protocol, or MCP, which is an open source standard for connecting models and agents to external tools, released by Anthropic at the end of 2024. It is thanks to MCP that Claude Code can, among other things, browse the web, manage files, interact with GitHub, deploy a container, and more, frequently completing tasks instantly.
However, powerful as MCP was, it had one weakness: the problem of context bloat.
Until now, Claude Code had to essentially load the whole “instruction manual” of every available tool during initialization, whether he used them or not. With more complex ecosystems of tools, such an approach took up an enormous amount of the context window of the model before the user typed a sentence.
This dynamic will shift now that there exists a new functionality called MCP Tool Search, which is actually an example of ‘lazy loading’ of AI tools.
The ‘Startup Tax’ on AI Agents
The challenge MCP was facing was scale. As more and more developers created MCP servers, each of these servers was providing a list of dozens, and sometimes hundreds, of tools.
"The problem, as anthropic engineer Thariq Shihipar succinctly put it, was that some MCP servers now housed 50+ tools, and that running setups involving 7+ servers was using 67,000+ tokens during startup," wrote Angela Chen, describing events that
To be exact, in extreme scenarios, a single Docker MCP server may use 125,000 tokens to simply describe its 135 tools.
In other words, it meant giving up a range of 30- to 50% of their context window prior to asking any kind of question. In explanation, according to an AI writer named Aakash Gupta, the models had to read through hundreds of pages of documentation they might never use.
But it was a hard trade off, because if you were to restrict your tools, it would badly impact agent capability; or if you wanted better reasoning, there would be heavy context loss:
How MCP Tool Search Fixes the Problem
MCP Tool Search brings in an intelligent approach.
Instead of preloading all definitions of tools, Claude Code watches how much context a description of tools would take up. If they go beyond 10% of context, it shifts modes.
Instead of spewing the raw documentation data into the question, Claude loads the light search index for the available tools. Upon the user request for an action, for instance "deploy the container," Claude will scan the index for the appropriate tool and load only the definition of that tool into the context.
The impact is very striking. Internal data released by Gupta specifies that token usage metrics have gone down from around 134,000 tokens to 5,000 tokens, a reduction of 85%, while still enjoying full access to all the same tools.
This also affects people who develop MCP servers that run on these tools. It changes best practices as follows: The server instructions field was optional; it became crucial. It has a function as metadata that informs Claude when it needs to search for the tool.
Better Focus, Better Accuracy
"The most obvious gain is token savings, but the secondary effect might well be even more significant: better reasoning quality."
The model is distraction friendly. The presence of hundreds of irrelevant definitions in the tool lexicon in the context window introduces the ‘needle in a haystack’ effect. The similar model commands have similar identifiers, which can impact decision-making.
Boris Cherny, head of Claude Code, reinforced this view after its release, saying that “users will get way more context, better instruction following, and the ability to plug in even more tools.”
Benchmarks confirm these findings. By community-shared results, accuracy of MCP evaluation with enabled Tool Search for Opus 4 increased from 49% to 74%, and for Opus 4.5, it increased from 79.5% to 88.1% accuracy.
In other words, by filtering out any kind of noise that is not relevant, the model will be able to focus on not only the request of the user but also on those tools that are relevant.
A Sign of a Maturing AI Stack
This announcement is a part of a wider shift in the development of AI.
Initially, early platforms use the principle of brute force. However, once they are scaled, architectural discipline is not optional but necessary. Gupta compared this to current software development tools. An IDE like VS Code or JetBrains doesn’t load all its plugins or extensions at the time of its initiation; it loads the required component whenever needed.
MCP Tool Search translates this concept to agents. Here, it uses dynamic, demand driven imports rather than static imports.
This is a significant indication. “AI systems are no longer research prototypes, but are instead emerging as sophisticated software platforms requiring a balance between capability, efficiency, and performance.”
What This Unlocks
For the user, there is little change at all. Claude Code simply feels faster, smarter, and more attentive.
For application developers, it's an entirely different ball game. This means that the soft cap on an agent's abilities is totally away. An agent can have possession of hundreds or even thousands of services, namely tools think databases, deployment scripts, APIs, file systems before an application has to pay for context.
According to the summary by Gupta, the shift alters the meaning of "tool rich agents." The context economy is from scarcity to access.
The update will be deployed immediately in Claude Code users, but it’s advised that developers of the MCP client integrate the new ToolSearchTool in order to support dynamic loading.
To summarize: Claude Code didn’t become more efficient; in fact, a huge leap towards a scalable agent system was made.
👉🏻 Found this article interesting? Follow us on Facebook, Twitter and whatsapp to read more exclusive content we post.
