Triage AWS costs with a slimmer MCP server
Author rebuilt an AWS cost-triage MCP server from 15 tools to 4, cutting a question from 7 tool calls to 2 with schemas 88% smaller.
Done with a custom-built or unnamed agent
- What they did
- The author built a read-only AWS cost-triage MCP server twice, once as 15 thin boto3 wrappers and once as 4 outcome-level tools. Both ran in Kiro against the same real account, asked where money was being wasted. The lean version has a find_waste tool that runs the checks internally and returns five fields per finding. It also has constrained parameters and a separate estimate_savings tool.
- How it went
- The bloated server needed 7 calls and the lean one 2. Tool descriptions shrank 88% (7,898 to 897 characters), and both found the same three wasted resources. The Elastic IP cost stayed unverified.
- Worth knowing
- On this small demo, Kiro's context meter only moved from 1.1% to 0.8%, so expect big savings only with many tools.
Try it yourself with Claude Code
Review my MCP server in [path to project] that exposes [number] tools for [e.g. AWS cost data]. Propose a slimmer design that merges overlapping tools into as few as possible, trims the schemas, and still answers questions like [example question]. Show me the plan and the before and after tool counts before editing any files, then implement it and compare the number of tool calls on that question.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Source: DEV · Oct 26, 2026
Related use cases
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Let a custom agent change its own repository
Author's custom-built agent made its first code change to its own repository.
Custom agent
Generate internal dashboards and data migration scripts
A backend engineer says their team uses Devin for internal dashboards and migration scripts, treating output as a starting point that saves about four hours of boilerplate.
Build and submit a family quest app to the App Store
A user built Kogoto Quest, which turns family reminders into quests, with Replit Agent and submitted it for App Store review.
Queue maintenance and bug fixes for Cursor background agents
A developer authorized the GitHub app so Cursor background agents cloned repos into sandboxes, branched, and worked through a day's queue of maintenance, features and bug fixes.
Build apps with a judge subagent and multi-stage review
A Claude Code setup for app development uses a judge subagent, context management, and multi-stage review for quality control.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.