I tested 3 Claude Code plugins to reduce costs. Here’s what actually worked

For context, I’m working on an internal CRM builder with a multi-agent setup (orchestrator + dev agents + reviewer that can iterate over the same task for many turns). When I looked into where most of the tokens were going, I realized it wasn’t the generated code, but the cache reads. And thinking about it, it makes sense... every turn resends the growing context with agent instructions, tool definitions, conversation history, etc. When you multiply that across several agents ping-ponging over a ticket, the token bill adds up fast. So I tested 3 plugins to see what could reduce either the context size or the number of turns (ranked from best to worst): Ponytail: It forces the agent through a checklist before coding, stopping once it finds the simplest solution. In my case, it cut generated code by around 50%, reduced costs significantly, and led to fewer dev/reviewer iterations. Definitely worth trying if you’re also working with code-generation loops or if you already know that your agents tend to over-engineer. Caveman: It tells the agent to remove filler and keep responses concise. It led to 65% fewer output tokens. However, unfortunately in my setup, the extra instructions added to every turn outweighed those savings. I think this could be pretty useful for conversational agents, less so for code-generation loops. Headroom: It compresses context before sending it to the model. Its token mode reduces cache reads by around 10%, but rewriting the history destroyed the existing cache prefix and triggered much more expensive cache writes. Its cache mode preserved the prefix, but effectively didn’t compress anything in our setup. This is a good example of why “fewer tokens” doesn’t necessarily mean “lower cost” when prompt caching is involved. I’m still looking to cut down on token usage in my multi-agent setup, so any tips are welcome! ( FYI I’m not affiliated with any of these plugins )