Claude Code: Tools Speed Writing and Code Workflows

CodeGraph provides a pre-indexed semantic code knowledge graph for code-focused LLM agents such as Claude Code, Cursor, Codex, and OpenCode, cutting tool calls and token use dramatically. In benchmarks across six real-world codebases (VS Code, Excalidraw, Claude Code itself, a Java project, Alamofire, and the Swift compiler) using Claude Opus 4.6 and Claude Code v2.1.91, CodeGraph reduced tool calls by ~92% on average, sped up exploration by ~71%, and eliminated file reads during agent runs. Agents query the graph (codegraph_explore) instead of repeatedly running grep/find/read, enabling faster, cheaper cross-language tracing and call-chain discovery; the Swift compiler (25k files) was indexed in under four minutes. This matters because it lowers token costs, improves latency, and enables scalable semantic code understanding for developer tooling and code-assistant workflows.

6.4k Stars！用Claude Code写论文的全套流水线，有人打包开源了

CodeGraph is an open-source, local tool that pre-indexes codebases into a knowledge graph so AI coding agents like Claude Code can query structure, symbols, call graphs, and imports in one call instead of repeatedly grepping files. Built with tree-sitter for parsing and SQLite+FTS5 for storage, it resolves references, auto-syncs via file watchers, and runs as an MCP server with MIT license and no external APIs. The maintainer reports benchmarks across several real projects showing 92% fewer tool calls and 71% faster exploration, with dramatic reductions in file reads for queries on VS Code and other repos (self-reported). Besides speeding agents, CodeGraph offers practical non-AI tooling such as an affected-tests tracer useful for CI. The benefit scales with repo size and familiarity.