The staleness cascade is the clever part: an edit only invalidates dependents when exports or types changed, so re-processing stays small and the index keeps up with you. Everything but the summaries works without any LLM at all, which makes the optional local model a genuine option rather than a prerequisite.
A codebase index that watches your repo, ranks files by importance, maps dependencies and keeps AI-generated summaries fresh in the background.
- find_important_files and list_files — top files by importance score with dependency counts, or the full tree
- get_file_summary — summary, concepts, change impact, exports, dependencies and staleness for one file
- find_symbol, find_callers and find_callees — resolve a name to a file and line range, and walk the TS/JS call graph both ways
- search across symbols, summaries, purpose and paths, and list_changed_since a timestamp or git SHA
- get_communities, detect_cycles and get_cycles_for_file — Louvain-clustered modules and circular dependency chains
- status for broker connection, queue depth, LLM progress and watcher state
- scan_all, set_base_directory, set_file_summary, set_file_importance and exclude_and_remove for managing the index
- Importance scores from 0 to 10, based on what depends on a file, what it exports and where it lives
- A Nexus dashboard at localhost:1234 with dependency graphs and per-repo health, auto-discovering every indexed repo on the machine
Node.js 22 or higher and npm, plus build tools for native modules — build-essential and python3 on Linux, Xcode command line tools on macOS, Visual Studio Build Tools with the C++ workload on Windows. Clone and run build.sh, which registers it with Claude Code. LLM summaries are optional and need llama-server or any OpenAI-compatible HTTP endpoint; without it, file tracking, dependencies, symbols and call graphs still work. Add .filescope/ to your .gitignore.