Large language model (LLM) agents usually rank each tool independently by request similarity, which works for single‑tool queries but misses prerequisite tools needed for multi‑tool workflows. LIBRA converts a flat catalog into an executable tool library. For every tool the LLM proposes a contract describing candidate arguments, reusable outputs, state changes and success conditions; LIBRA then executes these compositions in a sandbox and keeps the relations that succeed at runtime. The library is built before any user query, requiring no dependency annotations, usage trajectories or task‑specific fine‑tuning, and it supports both one‑shot and iterative agentic retrieval. Benchmarks on AppWorld, BFCL and SciToolEval show that LIBRA improves one‑shot recall and NDCG over the same dense retriever by an average of 7.9 % and 3.7 %. In the ToolReAGt setting, after ten retrieval iterations the gains are 4.5 and 2.9 points respectively.\n\nReview: By materializing tool dependencies through runtime testing, LIBRA markedly boosts retrieval performance for complex tasks, proving that LLMs can automatically construct executable tool libraries without extra supervision.