Semble uses 98% fewer tokens than grep for code search

Semble, a code search tool optimized for AI agents, uses 98% fewer tokens than grep while delivering equivalent results [GitHub]. The project, released May 17, 2026, targets agent-first workflows where token efficiency directly impacts speed and cost [Hacker News].

Unlike grep, which scans files linearly, Semble preprocesses codebases into a queryable format that reduces redundant data exposure during search. This design cuts token usage by compressing context and eliminating repetitive file headers, boilerplate, and comments from output. The tool supports regex patterns and integrates with existing CLI environments, allowing developers to swap grep with minimal friction.

Semble’s efficiency gains matter most in large repositories or automated systems. On a 100k-line codebase, grep typically outputs 15,000–20,000 tokens per search; Semble averages 300–500 [GitHub]. That reduction extends to agent loops where repeated searches compound costs.

The tool is already being tested in agent frameworks like LangChain and AutoGPT, where token savings translate to faster iteration and lower API bills. Its MIT license and lack of external dependencies make it easy to embed.

While grep remains embedded in countless scripts and muscle memory, Semble offers a drop-in alternative for modern, agent-driven development. Early benchmarks are limited to Linux x64 with glibc, but the team plans support for Windows and macOS [Hacker News].