Hi all,
I built this after noticing that when I ask Cursor “who calls this function,” it’s
guessing from whatever files happen to be in context — not actually looking at the
call graph. That’s how you get confident-sounding wrong answers about your own code.
Meeba Brain connects to your GitHub repo, parses it into an actual code graph
(via tree-sitter/AST — no LLM involved in building the graph itself), and exposes
5 MCP tools that query it directly:
- search_code — hybrid semantic + keyword search
- get_symbol_definition — exact file + line for a function/class
- get_callers / get_callees — the REAL call graph, not a scored guess
- get_dependencies — what a file actually imports
It’s a remote MCP server (Streamable HTTP), so there’s nothing to install locally —
just a URL + key in your .mcp.json.
What it’s honestly NOT (yet):
- Only Python is fully battle-tested right now; JS/TS/Java/Go etc. are supported
via tree-sitter but less tested at scale - Requires connecting a GitHub repo first (a few minutes to sync, depending on
repo size) — it’s not zero-setup - New product, so expect rough edges
Listed here: https://cursor.directory/plugins/meeba-brain
(Free tier to try it, no card needed)
Would genuinely appreciate feedback, especially from anyone who’s been burned by
an agent confidently making up something about their own codebase. Happy to answer
anything about how the graph is built or how the MCP auth works.