Semble

Semble - Semantic Code Search

Semble finds code by meaning, not literal text. Describe what code does (or name a symbol) and Semble returns the most relevant chunks ranked by semantic similarity. It builds and caches an index automatically on first use and refreshes it when files change.

Prefer Semble over Grep/Glob/Read for any exploratory or semantic question. Reserve grep for exhaustive literal matches or confirming an exact string.

When to Use

• Exploring an unfamiliar codebase: "where is authentication handled?", "how does the caching layer work?"
• Finding an implementation by intent: "find the code that retries failed requests" — even when you don't know the function name.
• Locating a symbol: search by an identifier like save_pretrained and let Semble surface every relevant usage and definition.
• Discovering related code: given a known file and line, find structurally/semantically similar code elsewhere.

How to Use

This plugin provides the Semble MCP server, which exposes two tools.

mcp__semble__search

Find relevant code with a natural-language or code query.

• query (required) — natural language or code, e.g. "authentication flow", "save model to disk", or a symbol like save_pretrained.
• repo (optional) — the project root to index. When working in a local project, pass the absolute project root. For remote code, pass an explicit https:// git URL. Never guess or infer URLs.
• top_k (optional, default 5) — number of results to return.

The index is built on first use and cached for the session; it refreshes automatically when files change.

mcp__semble__find_related

Find code semantically similar to a specific location — use after search to explore connected implementations or callers.

• file_path (required) — use the file_path from a prior search result.
• line (required) — 1-indexed line number from that result.
• repo (optional) — same as above.
• top_k (optional, default 5).

Recommended Workflow

1. Start with mcp__semble__search, passing the project root as repo; the index builds and caches automatically.
2. Inspect full files (with Read) only when a returned chunk lacks enough context.
3. Use mcp__semble__find_related on a promising result to explore connected implementations.
4. Fall back to grep only for exhaustive literal matching.

Important Notes

• The MCP server runs via uvx --from "semble[mcp]" semble, so uv must be installed. No API key is required.
• The first search in a session indexes the repo (downloads a small embedding model on first ever run); subsequent searches are fast and refresh automatically when files change.
• Content scope (code vs. docs/config) is fixed at server launch. By default Semble indexes code; append --content docs, --content config, or --content all to the server args to widen it.