A practical investigation into Claude Code's tool delegation when using third-party models reveals a crucial detail: all built-in tools—web search, image analysis, and web fetch—are fully delegated to the configured third-party model, not Anthropic's native capabilities. This means that if you configure Zhipu's GLM, the web search tool uses GLM's search, not Claude's. For developers, this has immediate implications: tool behavior may differ significantly between providers, affecting reliability and output quality. The post provides a clear testing methodology to verify this behavior, which is essential for anyone building workflows with Claude Code and alternative models. It also raises broader questions about tool portability and standardization across LLM providers. This is a timely signal for the AI engineering community as multi-model setups become more common.
This post investigates a critical question for Claude Code users: when you configure a third-party model like Zhipu's GLM, do the built-in tools (web search, image analysis, web fetch) still use Anthropic's native capabilities or are they replaced by the third-party model's equivalents? The author's testing reveals that all tools are delegated to the configured third-party model. This has significant implications for developers relying on specific tool behaviors and highlights the importance of understanding model-tool compatibility in multi-provider setups.