Ontomics

What EtienneChollet/ontomics MCP server does

EtienneChollet/ontomics MCP server gives coding agents a semantic view of a local repository. Instead of limiting queries to exact text matches or symbol references, it indexes domain concepts, naming patterns, code entities, relationships, type flows, and function behavior. The supported source languages are Python, TypeScript, JavaScript, and Rust, selected automatically from file extensions.

The index can help answer questions such as which concepts recur across a project, how a concept moves through call chains, which names fit established conventions, and which functions have similar implementations even when their names differ. It can also compare the indexed vocabulary with a Git reference and export domain knowledge as YAML for reuse in another repository.

How it works

The server runs its analysis locally and does not require API keys. Tree-sitter extracts identifiers, signatures, and call sites. TF-IDF then highlights domain-specific terms and naming patterns. BGE-small creates concept embeddings, while CodeRankEmbed creates 768-dimensional embeddings for function bodies and supports behavioral clustering. PageRank contributes structural-importance scores for entities.

Embedding models download on the first run and are cached locally. The resulting database is stored at <repo>/.ontomics/index.db. Later starts use the cached index and watch for file changes. A repository briefing is also available through the ontomics://briefing resource and the ontomics briefing command.

Setup and configuration

Install the binary through the published npm package, Homebrew, a shell installer, or a source build. The npm package is @ontomics/ontomics. For a project-level MCP configuration, the README shows a stdio server using npx -y @ontomics/ontomics --repo ..

The server normally detects the repository automatically and requires no configuration. It expects a Git repository containing a .git/ directory and rejects home, root, and temporary directories. Use --force when indexing a non-Git directory. Project-specific settings belong in .ontomics/config.toml; defaults are provided for all fields.

Tools and capabilities

EtienneChollet/ontomics MCP server exposes tools in four groups:

• Concept and vocabulary analysis: query, locate, describe, trace, and list concepts; inspect conventions; list entities; check or suggest names; measure vocabulary health; compare ontology changes; and export a domain pack.
• Behavioral analysis: find similar logic, describe a function's logic, and assemble compact context for a symbol.
• Codebase structure: describe files, map concepts to modules, inspect type flows, and trace types across files and call sites.
• Session context: retrieve the ontomics://briefing resource with conventions, abbreviations, key concepts, contrastive pairs, and vocabulary warnings.

Limitations and notes

EtienneChollet/ontomics MCP server is intended for local codebase analysis rather than hosted repository access. The provided material documents macOS and Linux support, not Windows support. Only Python, TypeScript, JavaScript, and Rust are listed as supported languages.

Initial indexing includes downloading the embedding models, so the first run differs from later cached starts. Non-Git directories require the explicit --force option. The benchmark reports one-call, lower-token results for selected questions, but those measurements are examples from specific codebases and do not establish performance for every project.