Matt Linderman retweetledi

Anthropic just solved MCP’s biggest lie: that 50+ tools means 50+ capabilities.
The math tells a different story. Users with 7 MCP servers were burning 67k tokens before typing a single character. That’s 33% of Sonnet’s 200k context window gone. One Docker MCP alone consumed 125k tokens with 135 tools.
Tool Search flips the architecture. Instead of preloading every tool definition upfront, Claude Code now loads a search index and pulls tools on-demand. The token savings are dramatic: from ~134k to ~5k in Anthropic’s internal testing. That’s an 85% reduction while maintaining full tool access.
But here’s what makes this actually interesting. They’re not just optimizing context usage. They’re changing what “tool-rich agents” can mean.
The old constraint forced a brutal tradeoff: either limit your MCP servers to 2-3 core tools, or accept that half your context budget disappears before you start working. Power users documented setups with 7+ servers consuming 67k+ tokens. The GitHub issue tracking this had 97 upvotes. People were building CLI tools just to toggle servers on and off between sessions.
Tool Search reframes the problem. The bottleneck wasn’t “too many tools.” It was loading tool definitions like 2020-era static imports instead of 2024-era lazy loading. Every IDE figured this out years ago. VSCode doesn’t load every extension at startup. JetBrains doesn’t inject every plugin’s docs into memory. Claude Code was architecturally behind its own ecosystem’s best practices.
The accuracy improvement matters more than the token savings. Opus 4 went from 49% to 74% on MCP evals with Tool Search enabled. Opus 4.5 from 79.5% to 88.1%. Fewer tools in context means less confusion between similar tool names like notification-send-user vs notification-send-channel.
This is Anthropic doing infrastructure work that doesn’t generate headlines but compounds over time. The MCP ecosystem just got permission to build bigger.
Thariq@trq212
English










