Semble MCP Server

Content

Semble is a local-first code-search MCP server for coding agents. It indexes local paths or remote git repositories and exposes focused search results through MCP instead of making Claude grep broadly and read entire files.

The MCP server provides search for natural-language or code queries and find_related for finding chunks similar to a known file and line. The project uses CPU-only hybrid retrieval with static code embeddings, BM25, tree-sitter chunking, and reranking. Indexes are cached and local repositories can be watched for automatic re-indexing during an MCP session.

Source Review

• https://github.com/MinishLab/semble
• https://github.com/MinishLab/semble/blob/main/README.md
• https://github.com/MinishLab/semble/blob/main/docs/installation.md
• https://github.com/MinishLab/semble/blob/main/pyproject.toml
• https://github.com/MinishLab/semble/blob/main/src/semble/mcp.py
• https://github.com/MinishLab/semble/blob/main/src/semble/cache.py
• https://github.com/MinishLab/semble/blob/main/benchmarks/README.md
• https://github.com/MinishLab/semble/blob/main/LICENSE
• https://pypi.org/pypi/semble/json

These sources were reviewed on 2026-06-05. Prefer the live repository, README, installation guide, Python project metadata, MCP server implementation, cache implementation, benchmark notes, license, and PyPI metadata for current commands, package versions, supported clients, tool behavior, cache behavior, and benchmark claims.

Features

• Search local repositories by natural-language or code query.
• Search remote git repositories by passing an explicit repository URL.
• Find code chunks related to a file path and line number from a prior result.
• Cache indexes on disk for faster repeat queries.
• Watch local paths in MCP mode and rebuild indexes after file changes.
• Index code by default, with optional documentation, configuration, or all-file scopes.
• Use .gitignore and .sembleignore rules to control indexed files.
• Install MCP configs, usage instructions, or a dedicated search sub-agent for supported clients.
• Run on CPU without requiring API keys, GPUs, or external vector services.

Installation

The recommended installer uses uv:

uv tool install semble
semble install

For manual MCP setup, configure the server to launch through uvx:

{
  "mcpServers": {
    "semble": {
      "command": "uvx",
      "args": ["--from", "semble[mcp]", "semble"]
    }
  }
}

To include documentation, configuration, or every supported file type, append --content docs, --content config, or --content all to the server command.

Use Cases

• Ask Claude where a feature is implemented without loading every matched file.
• Search unfamiliar repositories by behavior, API name, or architecture concept.
• Find related implementations after inspecting one relevant search result.
• Let a coding agent query documentation and config only when that scope is approved.
• Add AGENTS.md or CLAUDE.md guidance so agents prefer Semble before broad grep/read.
• Install a dedicated semble-search sub-agent in harnesses that support sub-agents.

Safety and Privacy

Semble is intentionally close to source code. It can index local repositories, clone or index remote git repositories, persist cache data, and return source chunks to the MCP client. Scope each search to repositories you are authorized to expose to an AI assistant.

Use .gitignore and .sembleignore to exclude generated files, vendored code, secrets, private fixtures, and unrelated projects. Treat caches and usage statistics as source-derived artifacts, especially when using documentation, configuration, or all-file indexing modes.

Duplicate Check

No MinishLab/semble entry, semble PyPI package entry, or matching source URL was found in content/mcp.