#2424th of 52 in MCP servers for developers

semble

Local code search for coding agents, via MCP server or CLI

Stars
6.3k
License
MIT
Last commit
Oct 2026
Last release
Oct 2026
Language
Python
Stars, last 28 weeks+6,252
7 Apr 2026: 3 stars14 Apr 2026: 3 stars21 Apr 2026: 5 stars28 Apr 2026: 202 stars5 May 2026: 672 stars12 May 2026: 788 stars19 May 2026: 3,082 stars26 May 2026: 4,302 stars2 Jun 2026: 4,663 stars9 Jun 2026: 4,931 stars16 Jun 2026: 5,126 stars23 Jun 2026: 5,295 stars30 Jun 2026: 5,380 stars7 Jul 2026: 5,468 stars14 Jul 2026: 5,545 stars21 Jul 2026: 5,604 stars28 Jul 2026: 5,668 stars4 Aug 2026: 5,766 stars11 Aug 2026: 5,817 stars18 Aug 2026: 5,860 stars25 Aug 2026: 5,914 stars1 Sept 2026: 5,946 stars8 Sept 2026: 5,996 stars15 Sept 2026: 6,052 stars22 Sept 2026: 6,105 stars29 Sept 2026: 6,142 stars6 Oct 2026: 6,172 stars13 Oct 2026: 6,198 stars20 Oct 2026: 6,255 stars

3 on 7 Apr 20266,255 on 20 Oct 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

Semble indexes a local path or git URL with tree-sitter chunking, then ranks chunks by fusing static Model2Vec embeddings (potion-code-16M-v2) with BM25 using Reciprocal Rank Fusion. It runs on CPU and returns only matching snippets to agents such as Claude Code, Cursor, Codex and OpenCode. It ships as an MCP server with search and find_related tools, a CLI, a sub-agent installer and a Python library.

Who it is for: Developers giving Claude Code, Cursor or Codex fast local code search

Strengths

  • Runs on CPU with no API keys or external services; model downloads once from Hugging Face
  • Reports NDCG@10 of 0.855 over ~1,250 queries across 63 repos and 19 languages
  • Searches several repos or git URLs as one corpus with incremental, cached indexes
  • Interactive `semble install` configures MCP, AGENTS.md instructions or a sub-agent per detected agent

Weaknesses

  • First run needs network access to fetch the embedding model unless SEMBLE_MODEL_NAME points locally
  • Files over 1 MB and data files such as .json and .csv are skipped by default
  • The ~99% token savings figure is the project's own benchmark against grep+read
  • Retrieval only; it does not edit code or run commands

What it needs

  • no GPU
  • Needs uv, Hugging Face (first-run model download)
  • Models: Model2Vec static embeddings (potion-code-16M-v2)

Also in MCP servers for developers

See all 52
Also in MCP servers for developers
RankProjectScore
1gitnexusCode knowledge graph exposed to coding agents through MCP tools Live demo ↗ (opens in a new tab)48k starscustom license92 out of 100
2codegraphLocal code knowledge graph that coding agents query over MCP74.6k starsMIT82 out of 100
3n8n-mcpMCP server exposing n8n node docs, templates and workflow management to AI assistants Live demo ↗ (opens in a new tab)23.1k starsMIT77 out of 100
#4serenaMCP server giving coding agents symbol-level code search, editing and refactoring30.5k starscustom license74 out of 100
#5chrome-devtools-mcpMCP server that lets coding agents drive and debug a live Chrome53.9k starsApache-2.072 out of 100