embecode

Local-first MCP server for semantic + keyword hybrid code search. Zero external services. No API keys required.

Usage

# From your project root
uvx embecode

# Or with an explicit path
uvx embecode --path /path/to/repo

Add to your MCP client config (Claude Desktop, Cursor, Cline, etc.):

{
  "mcpServers": {
    "embecode": {
      "command": "uvx",
      "args": ["embecode"]
    }
  }
}

Related MCP server: Code Memory

Tools

How it works

• Parses files into AST chunks via tree-sitter (cAST algorithm)

• Embeds chunks locally with sentence-transformers (nomic-embed-text-v1.5)

• Stores vectors + FTS index in a single DuckDB file at ~/.cache/embecode/

• Fuses BM25 and dense vector results with Reciprocal Rank Fusion

• Watches for file changes via watchfiles and re-indexes incrementally

Development

# Install dependencies
uv sync

# Run tests
uv run pytest

# Lint and format
uv run ruff check src/ tests/
uv run ruff format src/ tests/

Benchmarks

Two benchmark classes live in tests/test_performance.py and use pytest-benchmark:

Run the real benchmarks:

pytest tests/test_performance.py::TestSearchBenchmarkReal -v --benchmark-only --no-cov -s

The first run builds a 200-file synthetic index into .bench_db/ (~20s). Subsequent runs reuse it and start immediately. Delete .bench_db/ to force a rebuild.

Run the mock benchmarks (no setup cost, useful for isolating Searcher logic overhead):

pytest tests/test_performance.py::TestSearchBenchmark -v --benchmark-only --no-cov -s

Reading the output:

Each test prints a per-phase timing breakdown from SearchTimings on the last benchmark round:

phase breakdown (last run): {'embedding_ms': 0.0, 'vector_search_ms': 78.5, 'bm25_search_ms': 6.5, 'fusion_ms': 0.01, 'total_ms': 85.0}

pytest-benchmark then prints a summary table with min, max, mean, median, and stddev across all rounds.

Requires Python 3.12.