Stellarion provides a locally hosted MCP Server that connects AI clients and editors to a full‑code‑intelligence engine via the LSP protocol. By using on‑device tree‑sitter parsers, a semantic graph, and hybrid BM25‑vector retrieval, it delivers token‑efficient context selection that cuts AI development costs by 60–90% while offering real‑time code quality metrics and on‑prem security. The solution supports multiple languages, including JavaScript, Python, Go, and Rust, and runs on Windows, Linux, and macOS.
Funding
Funding not disclosed
Founders
Product
Problem
AI-assisted coding tools often require sending large amounts of source code to cloud services, leading to high token usage, increased costs, and security concerns. Developers also lack built‑in metrics to assess code quality, making it difficult to trust AI‑generated changes in large codebases.
Solution
Stellarion offers an on‑premise server that connects AI coding assistants and editors to a full code‑intelligence engine via the MCP and LSP protocols. The server builds a persistent semantic graph of the entire repository and provides a hybrid BM25‑vector search engine for token‑efficient context retrieval, reducing AI token consumption by 60–90%. Real‑time code quality metrics—including cyclomatic and cognitive complexity, maintainability index, and technical debt detection—are generated locally, enabling developers to evaluate the impact of AI suggestions before merging. All processing, indexing, and search occur on the user’s machine, ensuring data never leaves the premises and supporting compliance requirements such as HIPAA and SOC2.
Target Audience
Primary users are software development teams and engineering organizations that integrate AI coding assistants into their workflows and require on‑premise security, cost control, and code quality insights.
Features
- Hybrid BM25 + vector retrieval engine with sub‑millisecond similarity search using an HNSW index
- Persistent semantic graph storage (RocksDB) capturing import, call, and instantiation relationships across 17 tree‑sitter parsers (JS, TS, Python, Go, Rust, Java, C#, etc.)
- Real‑time code quality analysis delivering cyclomatic & cognitive complexity scores, maintainability index, dependency analysis, and impact assessment
- MCP and LSP protocol support for seamless integration with AI clients (Claude Code, Cursor, Windsurf) and IDE extensions (VS Code)
- On‑device embedding model (BGE) for zero‑API, fully local semantic search
- Token‑efficient context selection that sends only the most relevant code fragments to AI assistants