Semble – Code Search for AI Agents

TL;DR

TL;DR: Semble is an open-source code search tool that lets AI agents query any codebase in natural language, returning only the relevant snippets at 98% lower token cost than grep-plus-read — and it runs entirely on CPU with no API keys.

What Is Semble?

Semble is a code search library built specifically for AI coding agents. When an agent asks “How is authentication handled?”, Semble returns only the relevant code chunks — not full files — without burning through a 100k-token context window.

The core claim is backed by benchmarks on ~1,250 queries across 63 repositories in 19 languages: Semble hits 94% recall at around 2k tokens, while grep-plus-read needs the full 100k context window to reach 85% recall.

Source description (from the README): “Fast and Accurate Code Search for Agents. Uses ~98% fewer tokens than grep+read.”

Semble indexes a codebase in approximately 250ms and answers queries in approximately 1.5ms, all on CPU. The project is MIT-licensed, has 5,580 GitHub stars, and its latest release is v0.5.0 (published 2026-07-08).

Setup Workflow

Prerequisites

• Python 3.10+
• uv as the package manager

Step 1: Install Semble

uv tool install semble

Step 2: Run the Interactive Installer

semble install

The installer detects installed coding agents (Claude Code, Codex, OpenCode, Cursor, and others) and lets you choose which integrations to enable:

• MCP server — exposes Semble as a native MCP tool the agent calls directly.
• Instructions — adds CLI usage guidance to AGENTS.md or CLAUDE.md.
• Sub-agent — installs a dedicated semble-search sub-agent.

For fully unattended setup:

semble install --agent claude --type mcp subagent --yes

Step 3: Use It from Your Agent

Once installed, your agent queries Semble in natural language:

"How is authentication handled?"

Semble returns the relevant snippets with file paths and line numbers. No need to grep through files or read them in full.

Step 4: Update Semble

uv tool upgrade semble
uv cache clean semble    # for MCP users — restart MCP client after this

How It Works

Semble splits each file into code-aware chunks using tree-sitter, then scores every chunk against the query using two complementary retrievers:

1. Static Model2Vec embeddings — the code-specialized potion-code-16M model for semantic similarity. Because the embedding model is static with no transformer forward pass at query time, everything runs in milliseconds on CPU.
2. BM25 — for lexical matches on identifiers and API names.

The two score lists are fused with Reciprocal Rank Fusion (RRF), then reranked with code-aware signals including definition boosts, identifier stemming, and file coherence.

CLI Reference

Semble also ships as a standalone CLI for scripts or CI:

# Search a local repo (index built and cached automatically)
semble search "authentication flow" ./my-project

# Search a remote repo (cloned on demand)
semble search "save model to disk" https://github.com/MinishLab/model2vec

# Limit results
semble search "save model to disk" ./my-project --top-k 10

# Search docs or config instead of just code
semble search "deployment guide" ./my-project --content docs
semble search "config values" ./my-project --content config

# Find code similar to a known location
semble find-related src/auth.py 42 ./my-project

If semble is not on $PATH, use:

uvx --from "semble[mcp]" semble

Controlling Which Files Are Indexed

Semble respects .gitignore and .sembleignore (standard gitignore syntax, merged). .sembleignore lets you add rules without touching .gitignore:

# .sembleignore
generated/     # exclude generated directory
*.pb.go        # exclude Go protobuf files
!*.proto       # force-include Protobuf files

Well-known non-source directories (node_modules/, .venv/, dist/, build/, __pycache__/, etc.) are always skipped regardless of ignore files.

Practical Evaluation Checklist

• Indexing speed on a ~50k-line codebase (should be under 1 second)
• First-query latency after cold index
• Quality of results for ambiguous natural-language queries
• Token savings reported by semble savings
• Correctness of line number references in results
• Behavior when no results match the query
• Integration with your specific coding agent (Claude Code, Cursor, Codex, etc.)

Security Notes

• All processing is local — no data leaves the machine.
• No API keys or external services required.
• Remote git URLs are cloned on demand; no persistent remote connection.
• Cache is stored locally (~/.cache/semble/ on Linux, ~/Library/Caches/semble/ on macOS) and can be cleared with uv cache clean.

Source and Accuracy Notes

All claims in this post were verified from the following primary sources:

• Project page: github.com/MinishLab/semble
• README (read end-to-end): raw.githubusercontent.com/MinishLab/semble/main/README.md
• License: MIT (verified via github.com/MinishLab/semble/blob/main/LICENSE)
• Stars: 5,580 (verified via GitHub API, 2026-07-11)
• Latest release: v0.5.0 (verified via GitHub API releases endpoint, 2026-07-08)
• Benchmarks methodology: benchmarks/README.md on the repo
• Embedding model: huggingface.co/minishlab/potion-code-16M

FAQ

Q: Does Semble require a GPU? A: No. Everything runs on CPU. The static Model2Vec embedding model does not need a transformer forward pass at query time.

Q: Does it work with Claude Code, Cursor, and VS Code? A: Yes, via the MCP server integration. The README lists Claude Code, Cursor, Codex, OpenCode, and VS Code as supported.

Q: How does it compare to Greptile or other code search tools? A: Semble’s distinguishing claim is token efficiency: 98% fewer tokens than grep-plus-read at equivalent recall (94% vs 85%), with indexing at approximately 250ms per repo. The README acknowledges Greptile in its acknowledgements. Direct feature comparisons require running both on the same codebase.

Q: Can I use it as a Python library rather than CLI? A: Yes. The README includes a Python API example using SembleIndex.from_path() and index.search().

Q: What happens when code changes? A: In CLI mode, the cache is invalidated automatically when files change. In MCP mode, a file watcher detects changes and rebuilds the index within the same session.

Conclusion

Semble solves a real problem for AI coding agents: context window waste from reading full files just to find a few relevant lines. The benchmarks are compelling — 98% token reduction at higher recall than grep-plus-read — and the tool is production-ready with MCP support, a clean CLI, and MIT licensing. If your agent workflow involves large codebases, Semble is worth integrating today.