Seraph

Local code search and Git-aware code intelligence for coding agents. Seraph indexes a local Git checkout into a SQLite database and exposes search, symbol, dependency, history, and version comparison tools over MCP.

Install from a checkout

Requires Python 3.12 and Git. With uv:

uv tool install --from . seraph-mcp

For development directly from this directory:

uv sync
uv run seraph-mcp

The published package install, once released to PyPI, is:

python -m pip install seraph-mcp

Agent setup

Give your coding agent AGENT_SETUP.md. It explains how to install Seraph and register its local MCP server for that agent. MCP configuration differs by host, so the agent should use the host's own supported MCP setup mechanism rather than assume one universal config file.

Seraph runs on the user's machine. Each tool accepts repo, the absolute path to the local Git checkout to search. The agent should pass the active workspace path on every call; if asked about another checkout, pass that checkout's path. The index is stored at <repo>/.seraph/index.sqlite and ignored by Git.

MCP tools

• search_code: retrieve relevant code from the selected checkout.
• search_at_version: retrieve code from a commit, branch, or tag.
• search_history: search indexed Git versions.
• find_symbol: locate a function, class, or method.
• find_dependencies: walk call, import, and inheritance relationships.
• compare_versions: compare symbols, dependencies, and commits.
• index_repository: explicitly index a Git version; searches index automatically when needed.

The default retrieval setup is local and needs no API key. Benchmark/evaluation code and submission assets are intentionally excluded from this slim distribution.

Optional local semantic search

The default install uses BM25 and does not install Torch or download a model.

For BGE-Code-v1 or GTE-ModernBERT, opt in to the Transformers 4 extra:

python -m pip install 'seraph-mcp[cpu]'  # CPU / Apple Silicon
# or, on an NVIDIA CUDA machine:
python -m pip install 'seraph-mcp[gpu]'

Seraph downloads BGE-Code-v1 (about 6 GB) from Hugging Face the first time semantic search needs it. Hugging Face caches the files, so other repositories on that machine reuse the same model. Each teammate installs the extra and downloads their own local cache only if they opt in.

EmbeddingGemma-2 needs Transformers 5, so use its separate extra instead of cpu or gpu:

python -m pip install 'seraph-mcp[gemma2]'

Choose embeddinggemma-2 in Seraph's model configuration. The selected model weights download on first use and are then cached locally.

CLI

seraph --repo /path/to/repo index
seraph --repo /path/to/repo search "where is request authentication checked?"
seraph --repo /path/to/repo symbol authenticate_request
seraph --repo /path/to/repo deps authenticate_request
seraph --repo /path/to/repo compare <old-commit> <new-commit>

License

See LICENSE.