The thing most code servers stop short of is asking whether a path is dangerous. Here taint tracing, an OWASP and CWE rules engine, SBOM generation and OSV dependency checks sit next to the symbol and call-graph tools, so "who calls this" and "can user input reach it" are questions in the same session. Watch which flags you started it with, though: git, call graph, LSP and neural tools are each off by default, and the graph tools additionally need a source build.
A single ~30MB Rust binary that indexes your repositories with tree-sitter and exposes deep code understanding over MCP: symbol extraction and navigation, several kinds of search, control and data flow analysis, type inference, security scanning and supply-chain auditing. Nothing leaves your machine, and indexing runs in parallel across all cores.
- Symbol navigation across 32 languages — find definitions and references, analyse imports and dependents, fuzzy-search the workspace, read a module's export map — `find_symbols`, `get_symbol_definition`, `find_references`, `workspace_symbol_search`, `get_export_map`
- Four search strategies rather than one: keyword with relevance ranking, BM25, hybrid rank fusion, and similarity to a snippet or a symbol — `search_code`, `semantic_search`, `hybrid_search`, `find_similar_code`, `find_similar_to_symbol`
- Call graph analysis behind `--call-graph`: who calls a function, what it calls, the path between two functions, cyclomatic and cognitive complexity, and the most connected hotspots — `get_call_graph`, `get_callers`, `get_callees`, `find_call_path`, `get_complexity`, `get_function_hotspots`
- Control and data flow: basic blocks and branches, unreachable code, reaching definitions, variables used before initialisation, and assignments never read — `get_control_flow`, `find_dead_code`, `get_reaching_definitions`, `find_uninitialized`, `find_dead_stores`
- Type inference for Python, JavaScript and TypeScript without running mypy or tsc, including taint flow enhanced by the inferred types — `infer_types`, `check_type_errors`, `get_typed_taint_flow`
- Security analysis in two layers — taint tracking that traces user input to a sink, and a rules engine covering OWASP Top 10 2021, CWE Top 25, crypto and secrets — `find_injection_vulnerabilities`, `trace_taint`, `get_taint_sources`, `scan_security`, `check_owasp_top10`, `check_cwe_top25`, `explain_vulnerability`, `suggest_fix`
- Supply chain work: SBOM in CycloneDX, SPDX or JSON, known-vulnerability checks against the OSV database, licence compliance analysis, and safe upgrade paths — `generate_sbom`, `check_dependencies`, `check_licenses`, `find_upgrade_path`
- Git history behind `--git`: blame, file history, recent changes, churn-and-complexity hotspots, contributors, and the commits that changed a specific symbol — `get_blame`, `get_hotspots`, `get_symbol_history`, `get_modified_files`
- Neural semantic search behind `--neural`, which finds similar code even when the names differ, and detects Type-3/4 semantic clones — `neural_search`, `find_semantic_clones`
- Circular import detection and an incremental Merkle-tree change status — `get_import_graph`, `find_circular_imports`, `get_incremental_status`
A binary, from Homebrew, Scoop, `cargo install narsil-mcp`, `npm install -g narsil-mcp`, Nix, or the one-line install script. Building from source needs Rust 1.70 or later, and on Windows the Visual Studio Build Tools with C++ desktop development. Point it at repositories with `--repos`, repeated per repository. Several tool families are behind CLI flags — `--git`, `--call-graph`, `--lsp`, `--remote`, `--neural`, `--persist`, `--watch`, `--streaming` — and one is behind a build feature as well: `--graph` only works if the binary was built with `--features graph`, which the default binary is not, so SPARQL and Code Context Graph tools need a source build. Neural search with the `api` backend uses Voyage AI or OpenAI embeddings; the ONNX backend runs the model locally. Dual-licensed MIT or Apache-2.0.
One command — npm install -g narsil-mcp