serpapi-agent-toolkit

When you wire web search into a coding agent, the default MCP server floods the context window: every call drags in tokens you never asked for, and the model gets slower the more it searches.

serpapi-agent-toolkit is my answer to that. It is a self-authored, MIT-licensed command-line tool and Claude Code skill that gives an agent SerpApi search at a fraction of the cost. The URL builder and the result parser are pure functions with fetch injected, so all 37 tests run offline, with no network and no key. The judgment that matters here: I kept the surface small and compiled it to a single binary with bun build, instead of shipping one more always-on server for the model to reason about. CI runs the full suite on every push.

The result is about 17x fewer tokens per call than the default MCP, measured against real queries rather than estimated. That is the same operating discipline I bring to production client work: the core is pure functions with tests that run without the network, and the cost is a number I can point to.