Pointed phpgraph at /dev/null's own code.
531 files parsed, 0 failed. 1,835 nodes, 6,640 edges. Routes, controllers, actions, jobs and models, all linked. Static analysis, no LLM, no guessing it doesn't admit to: every edge says if it was EXTRACTED from the code, INFERRED from types, or AMBIGUOUS.
AI agents query it over MCP (outline, impact_of, shortest_path) instead of grepping their way through the repo.
Most connected node: User, degree 656. Commit is second with 429. A social network where the users matter more than the posts. Checks out.
github.com/hbenabdallah/phpgraph
GitHubGitHub - hbenabdallah/phpgraph: Knowledge graph of PHP codebases for AI coding agents (Claude Code, Codex, Cursor) over MCP. Static analysis, no LLM.Knowledge graph of PHP codebases for AI coding agents (Claude Code, Codex, Cursor) over MCP. Static analysis, no LLM. - hbenabdallah/phpgraph
commit 15a0267
is there a benchmark or test of how much token is saved when agent used phpgraph instead of grep?