codeseek

Semantic search over your codebase — as a CLI and an MCP server. Zero dependencies.

codeseek indexes a repository into a local vector store and lets you search it by meaning, not just by string match. Use it from the terminal, or run it as an MCP server so an AI coding assistant or editor can ask your codebase questions directly.

It brings no embedding model of its own: point it at the OpenAI API, or at any OpenAI-compatible endpoint such as a local llama.cpp server, so private code can be embedded without leaving your machine. Storage is plain SQLite; search is brute-force cosine. The whole thing is the Python standard library and nothing else.

Install

pip install codeseek
# or:
pipx install codeseek

Requires Python 3.8+.

Quick start

export OPENAI_API_KEY=sk-...

# 1. Index the current repository
codeseek index .

# 2. Search it
codeseek search "where do we validate the auth token?"
codeseek search "retry with backoff" -k 3

Results come back as markdown, each with a file path, line range, and similarity score:

### src/auth/token.py:40-70  (score 0.812)
​```
def verify_token(raw: str) -> Claims:
    ...
​```

Use a local or alternative provider

codeseek index . --base-url http://localhost:8080/v1 --model nomic-embed-text

As an MCP server

codeseek serve speaks MCP over stdio and exposes one tool, search_code.

After indexing a repo, register it with your MCP client. A typical mcpServers configuration looks like this:

{
  "mcpServers": {
    "codeseek": {
      "command": "codeseek",
      "args": ["serve", "--db", "/path/to/your/repo/.codeseek.db"],
      "env": { "OPENAI_API_KEY": "sk-..." }
    }
  }
}

The assistant can then call search_code to pull relevant code into its context on demand, instead of you pasting files by hand.

Commands

Shared options: --db, --model, --base-url, --api-key (each with an environment-variable default).

How it works

1. Source — files are walked and read (sensible code/text extensions, common build and vendor directories skipped).
2. Chunking — each file is split into overlapping line windows.
3. Embedding — chunks are embedded in batches via your provider.
4. Storage — vectors land in a local SQLite database.
5. Search — your query is embedded and compared against every chunk by cosine similarity; the top matches are returned.

The document source is pluggable: the engine only consumes Document objects, so the same indexing and search machinery can be pointed at things other than code.

Privacy note

Indexing sends file contents to whichever embeddings provider you configure. For private code, prefer a self-hosted model via --base-url.

Scope

Search is a linear scan, which is plenty fast for a single repository (a few thousand chunks). Indexing very large monorepos would want a real approximate vector index — a natural next step, not today's goal.

Development

pip install -e ".[test]"
python -m pytest

All tests run offline; the embedding and HTTP layers accept injectable fakes.

License

MIT — see LICENSE.