What it does

Semble is a code search tool built specifically for AI agents. You ask it a question in plain English—“How is authentication handled?"—and it returns only the relevant code snippets, without your agent grepping blindly or reading entire files. It works as an MCP server, a CLI, or a sub-agent template for Claude Code, Cursor, Codex, and others.

The interesting bit

The speed claims are the hook, but the integration strategy is the substance. Semble meets agents where they already live: MCP config files, AGENTS.md snippets, and harness-specific sub-agent directories. It also accepts git URLs, so an agent can search a remote repo it just cloned without any local setup beyond uvx.

Key highlights

• Indexes an average repo in ~250 ms, queries in ~1.5 ms, all CPU-bound with no API keys or GPU
• Claims NDCG@10 of 0.854, “on par with code-specialized transformer models” per its own benchmarks
• Supports natural-language queries, symbol search, and “find-related” from a specific file/line
• Can index code, docs, config, or all three via --content flags
• .sembleignore files let you tweak indexing without touching .gitignore

Caveats

• Benchmarks and token-savings figures are self-reported; the README links to them but the details are truncated in the provided source
• Requires uv for the smoothest MCP setup; pip works but the docs push uv heavily

Verdict

Worth a look if you’re burning context window on grep-and-read loops in agent workflows. Skip it if your agent already has fast, accurate codebase search built in.