If you've been using Claude Code for serious development work, you've probably noticed a frustrating pattern: every new session starts completely blind. Claude re-reads your entire codebase, rediscoveries architecture from scratch, and forgets every architectural decision you explained in the last session. Alex, a long-time CTO posting on Hacker News this week, built Graphmind to fix exactly that problem—and it's got some serious engineering under the hood.
The Core Problem Graphmind Solves
Claude Code's blank-slate approach works fine for quick scripts, but enterprise developers juggling multiple repositories quickly hit a wall. There's zero visibility into shared dependencies across projects. Pattern violations go unnoticed because Claude can't see what it did last week in another file. Graphmind attacks this with four distinct layers: a structural code graph built via AST analysis using tree-sitter, semantic embeddings for vector search over symbols (supporting local ONNX, OpenAI, or Voyage AI providers), a declarative memory store for decisions and conventions, and cross-project link tracking between registered repositories. Everything runs 100% locally—no cloud, no open ports by default, zero telemetry.
Architecture Deep Dive
The system is written in Rust using tree-sitter for multi-language parsing, with an MCP server exposing 24 tools over stdio. The structural graph lives in SQLite with FTS5 for full-text search, while embeddings use a separate SQLite database with cosine similarity and Reciprocal Rank Fusion (RRF) to merge results from multiple retrieval strategies. Search results show their source: [FTS], [SEM], [GRAPH], or combinations like [FTS+SEM+G]. The Rust core uses napi-rs for native bindings, and the MCP responses are deliberately token-optimized—compact one-line-per-symbol output by default, field pruning to strip null values, smart limits of 15 results instead of bloated defaults. That attention to LLM efficiency suggests someone who's actually shipped products that call these APIs.
Installation Options
Graphmind ships as both a CLI and macOS desktop app. The DMG includes guided onboarding that configures MCP, hooks, skill, and embeddings automatically—zero terminal required for Mac users. CLI installation works via Homebrew (macOS/Linux), raw shell script (curl piped to bash), or direct binary download for Linux. After a one-time graphmind setup, each project gets initialized with graphmind init which registers the directory, installs git hooks for auto-rebuild on commit and impact checking on push, then builds the code graph. Both commands are idempotent so they're safe to re-run. The desktop app even handles CLI updates automatically at startup.
Language Support and Data Storage
The tool supports 29 languages out of the box—TypeScript, Python, Go, Rust, Ruby, Java, C#, Swift, PHP, Kotlin, Scala, and many more are marked stable. All project data lives in ~/.graphmind/: structural graphs in SQLite under graphs/
Security Model
The architecture prioritizes local-only operation. MCP uses stdio transport (no network servers), path traversal is blocked to registered paths plus ~/.graphmind/, API keys stay in the config file and only go to their configured provider, memory writes use atomic tmp+rename operations, and the gm_memory_add tool requires explicit confirmation before storing anything. This isn't security theater—it's actual defense in depth for devs who don't want their code shipped off to some external service.
Cross-Project Intelligence
The feature that really caught my attention is cross-project link tracking. Register multiple repositories with graphmind init, and you get visibility into shared dependencies, common patterns across projects, and relationships between codebases. The graphmind cross query command searches symbols across all registered projects simultaneously. For developers maintaining microservices or shared libraries, this solves a real pain point that most IDEs still handle poorly.
Key Takeaways
- Four-layer architecture: structural graph + semantic embeddings + memory store + cross-project links
- 24 MCP tools for deep Claude Code integration with token-optimized responses
- Local-first design with zero telemetry—no code ever leaves your machine by default
- Supports 29 programming languages via tree-sitter AST parsing
- Available as CLI (Homebrew, shell install) or macOS desktop app with automatic updates
The Bottom Line
Graphmind is the kind of tool that makes you wonder why it didn't exist sooner. Alex built something genuinely useful for teams doing serious development work—not a toy demo but production-grade infrastructure for making AI coding assistants actually intelligent about your codebase over time. If you're running Claude Code professionally, this is worth installing today.