claude-context

Converts entire codebases into vector embeddings using pluggable embedding providers (OpenAI, VoyageAI, Gemini, Ollama) and stores them in a vector database (Milvus or Zilliz Cloud), enabling AI agents to retrieve semantically relevant code snippets without loading entire directories. Uses tree-sitter AST parsing for syntax-aware chunking across 40+ languages, with LangChain fallback for unsupported syntax.

Exposes semantic code search as a Model Context Protocol (MCP) server with standardized tool handlers, enabling Claude Code, Cursor, and other MCP-compatible AI assistants to invoke code search as a native capability without custom integration code. Implements MCP protocol with schema-based function calling and multi-project context management through a unified tool registry.

Tracks embedding generation costs, latency, and token usage per provider, providing visibility into indexing expenses and performance. Implements per-provider metrics collection with aggregation by time period and project, enabling cost optimization and provider comparison.

Manages system configuration through environment variables, configuration files, and CLI arguments with hierarchical precedence. Supports configuration validation, schema enforcement, and runtime configuration updates without server restart for non-critical settings.

Parses source code using tree-sitter AST parser to identify syntactic boundaries (functions, classes, modules) and chunks code at semantic boundaries rather than fixed line counts. Falls back to LangChain token-based splitting for unsupported languages, preserving code structure and enabling more precise semantic embeddings. Supports 40+ programming languages with language-specific chunking strategies.

Monitors filesystem changes using file watchers and Merkle-tree based change detection to identify modified files, avoiding full codebase re-indexing on every change. Implements delta-based synchronization that only re-embeds changed files and updates vector database entries, reducing indexing latency from minutes to seconds for typical code changes.

Abstracts embedding generation behind a provider interface supporting OpenAI, VoyageAI, Gemini, and local Ollama, allowing users to swap embedding models without code changes. Implements provider-specific batching, rate limiting, and fallback strategies, with cost tracking and performance metrics per provider.

Provides VS Code integration exposing semantic code search through IDE commands and UI panels, enabling developers to search their codebase without leaving the editor. Integrates with the core indexing engine and MCP server, displaying search results with syntax highlighting, file navigation, and one-click code navigation.

Manages multiple indexed codebases (projects) within a single MCP server instance, allowing AI agents to switch between projects and search across different codebases. Implements project-scoped vector database collections and configuration management to isolate indexes and prevent cross-project contamination.

Captures snapshots of vector database state at specific points in time, enabling rollback to previous index versions if embeddings become corrupted or outdated. Implements snapshot metadata tracking with timestamps and configuration checksums, allowing users to restore to known-good states without full re-indexing.

Supports configurable file filtering using glob patterns and gitignore-style rules to exclude files from indexing (node_modules, build artifacts, test files, etc.). Implements efficient pattern matching with early termination to avoid parsing excluded files, reducing indexing time and vector database size.

Provides browser extension for GitHub that enables semantic code search on public repositories without cloning. Indexes code directly from GitHub's web interface and stores embeddings locally or in cloud backend, enabling semantic search on any public repository.