Serena
Use the language server the IDE already runs, instead of building a search index
Search over a codebase answers “which files mention payment”. It doesn’t answer “what calls processRefund, and do any of those callers skip the audit log”.
Embedding search can’t answer the second one at all. It matches on words. A caller that reaches the function through getattr never mentions its name, so it never comes back in the results.
Every language server already knows the answer. Pyright, rust-analyzer and typescript-language-server all keep a live symbol graph, because that’s how your editor does jump-to-definition. So instead of building an index, this wraps the one that’s already running.
Serena is Oraios’ project. I extended it with an MCP interface for the team’s own repositories, using multilspy to talk to the language server:
from multilspy import SyncLanguageServer from multilspy.multilspy_config import MultilspyConfig, Language async def find_all_references( repo_path: str, file_path: str, line: int, character: int ) -> list[Location]: config = MultilspyConfig.from_dict({"code_language": Language.PYTHON}) async with SyncLanguageServer.create(config, repo_path) as lsp: return await lsp.request_references(file_path, line, character) That returns every call site across the repository in under 200ms, because the language server is already holding the index in memory.
Thirty tools are exposed to the assistant. The ones that get used: find every call site, jump to a definition across files, list what a file exports, get the call tree in both directions, and read the current errors.
What it changed
Reviewers can check what a change actually reaches before approving it, which matters most on paths that write to an audit log.
New engineers start with a session asking what a function does, where it’s called, and everywhere it touches the database. That used to be a few days of reading.
The team reckoned review cycles halved.