Your AI wrote the code; Noodles draws the map

What it does

Noodles ingests a repository or GitHub PR and uses tree-sitter to build a static function call graph. It renders the result as an interactive viewer you can pan, zoom, and drill into, plus Mermaid diagrams for quick sharing. An optional LLM pass—Anthropic by default, with OpenAI, Gemini, Groq, and HuggingFace supported—adds human-readable labels to nodes and edges; without an API key you still get the graph, just not the prose.

The interesting bit

The PR analyzer is the standout feature: it prunes the call graph to only functions touched by a pull request, turning a wall of AI-generated diff into a navigable map of what changed and why it matters. There is also an MCP server mode that plugs directly into Claude Code, letting you query callers, callees, and paths without leaving your editor.

Key highlights

• Parses Python, JavaScript, TypeScript, and TSX/JSX via tree-sitter; React component hierarchies show up as actual call edges.
• Runs fully offline for graph generation; only the descriptive enrichment step phones home to an LLM.
• Ships both a local interactive viewer and a hosted version at unslop.xyz.
• Integrates as an MCP server with tools like find_path, get_callers, and filter_graph for editor-native exploration.

Caveats

• Dynamic dispatch and framework callbacks—think Flask decorators or getattr chains—are invisible to the AST parser, so the graph can miss critical connections.
• The PR analyzer produces thin results when changes are isolated standalone functions with no callers or callees in the detected graph.
• Other languages require adding new tree-sitter grammars manually; only Python and the JS/TS family are supported today.

Verdict

Grab it if you are inheriting or reviewing sprawling AI-generated Python or JavaScript and need to orient yourself before refactoring. Skip it if your codebase relies heavily on dynamic metaprogramming or lives outside the supported language set.