Here’s something nobody warns you about when you give agents a web search tool: they use it constantly, and most of what they ask is trivial. Not wrong — just cheap questions. What’s the current date. What’s the name of that library again. Does this product still exist. Is this domain live. Little factual nudges that a search engine answers in a blink.
Every one of those goes out through a paid search API in the default setup, and the meter runs per query. I don’t have a precise number for what a routine agent lookup costs, but the shape of the problem is obvious the moment you watch an agent work for an afternoon. You’re paying a premium price for a dumb question, over and over, because it’s the only door you built.
So I built a second door.
The free lane is a metasearch engine. I run a self-hosted metasearch instance — metasearch meaning one search box that asks a pile of search engines at once and strips out the ads and trackers on the way back. Mine fans a single query out across 84 engines and 31 categories, no API key, no rate limit, no per-query charge. It sits on my own server, bound to the private network only, and my agents reach it the same way they reach everything else. Now every agent in the fleet has unlimited search for the routine stuff, and I’m not paying anyone by the question.
The paid lane handles the questions that actually deserve a specialist. A paid search API is genuinely better at three things. It returns AI-ready structured results instead of a page of links. It does semantic search — finding things by meaning when the exact words aren’t there. And it does the deep-research passes, the ones that chase a question through twenty sources and hand back a cited brief. My rule is simple: routine lookups go free, and paid credits get spent only when the free lane can’t answer.
That distinction did more for my bill than any discount ever has. I’d been treating search as one undifferentiated tool, so every question cost the same. Splitting it into two tiers meant the expensive lane stopped paying for the cheap work. Roughly seven or eight of every ten searches an agent makes turn out to be routine — that’s the share my free baseline now absorbs, and the paid API is reserved for the rest.
Here’s the honest caveat: sometimes you should just pay. Self-hosting isn’t free, it’s prepaid in a different currency. It went up in one sitting, and I carry the maintenance forever — the container needs monitoring, and I still owe myself a proper quality review after a week of real use to see whether specific query types need tuning. If you make a handful of searches a week, that trade is bad. Buy the API, keep your evening. This only wins when the volume is steady and the questions are mostly boring.
The free lane also isn’t a privacy cloak, in case that’s what you’re hoping. My server still reaches the search engines under my own address, and one person’s server looks the same every time. For an agent looking things up programmatically, that’s a non-issue — it’s what every paid search provider does from their own machines. For something personal, it isn’t.
What I actually bought was a routing habit. Not a cheaper tool — a decision about which question goes where, made before the query leaves the house.
Route your spend by question difficulty, not by habit.