What Devscriptor Does
for AI-Assisted Development
A toolkit built in Rust for turning your codebase into an AI-navigable knowledge graph
Built in Rust
A memory-safe engine with zero-cost abstractions, designed to run efficiently on everyday developer hardware.
- •Multi-threaded parsing (scales to num_cpus)
- •Memory-mapped file reading for zero-copy access
- •Batched, WAL-mode SQLite transactions
- •Content-hash caches skip redundant parsing & inference
Single SQLite Database
All your codebase knowledge in one portable file. Query with SQL, run FTS5 keyword and vector search, and maintain a persistent knowledge graph.
- •Portable single-file database
- •Directly queryable by any AI agent
- •FTS5 full-text search (code-aware tokenizer)
- •768-dim vector search via sqlite-vec
Code Graph Analysis
Comprehensive code analysis with entity extraction, relationship mapping, cross-reference detection, and architectural pattern recognition.
- •Entity extraction (functions, classes, variables)
- •Relationship mapping (calls, inheritance, imports)
- •Cross-reference detection across files
- •Architectural pattern recognition
LSP Code Intelligence
38 LSP tools providing real-time IDE features through a 4-tier detection system supporting 21 language servers.
- •Go-to-definition navigation
- •Hover information and type signatures
- •Code completions
- •Call and type hierarchy
- •Diagnostics and error detection
26 Languages & Frameworks
Tree-sitter parsing across systems, web, JVM, .NET, and mobile stacks — plus the Vue and Svelte frameworks — all with dedicated grammar support.
- •26 tree-sitter grammars (Tree-sitter 0.25)
- •TypeScript grammar covers both TS and TSX
- •Functions, classes, structs, enums, components…
- •Calls, inherits, implements & cross-file relationships
Gitignore-Aware Discovery
Parallel, multi-threaded file discovery using ripgrep's ignore engine, with native .gitignore support.
- •ignore crate (same engine as ripgrep)
- •All glob and negation patterns
- •Nested .gitignore files with proper precedence
- •.git/info/exclude support
Incremental & Real-time
Re-scans only what changed via an mtime+size fast path, and a watcher keeps the graph current as you code.
- •mtime+size fast-path, XXH3 content hashing
- •Only changed files are re-parsed
- •devscriptor watch for automatic updates
- •Content-hash embedding cache skips re-inference
Context Graph Memory
AI agents that remember. Store decisions, standards, and project knowledge across sessions with 9 structured note types.
- •9 note types (adr, standard, optimization, failure, practice, reflection, general, file_description, modification)
- •Notes live in the same local SQLite database
- •Linked to code entities and abstract concepts
- •Keyword, semantic & RRF hybrid search
100% Local & Private
Your code never leaves your machine. Scanning, indexing, and search make no network calls — there is no Devscriptor backend.
- •No code sent to external services
- •No telemetry or analytics
- •One-time embedding-model download, then fully offline
- •Pre-stage the model cache for air-gapped use
MCP Server Built-in
Native integration with Claude Desktop, Cursor, Cline, and any MCP-compatible AI agent. 75 total tools including 32 Core + 38 LSP + 5 Architecture.
- •32 Core analysis, context & watcher tools
- •38 LSP code intelligence tools
- •5 Architecture management tools
- •Unified server over stdio (devscriptor mcp)
Semantic Search
Find code by meaning, not just keywords. Local vector embeddings, no API keys, hybrid-ranked with keyword search.
- •Local jina-embeddings-v2-base-code (768-dim)
- •No API keys — runs offline once cached
- •Cosine similarity via sqlite-vec
- •RRF hybrid fusion of keyword + vector results
Relationship Mapping
Understand how your code connects. Automatic detection of inheritance, implementation, function calls, imports, and data flow.
- •Function call relationships
- •Inheritance and implementation
- •Import and dependency tracking
- •Cross-file relationship detection