Open-source Claude Code plugin routes grunt work to cheap models, saves Opus for judgment
investigatormaker · reddit · 2026-10-19
A Reddit user released claude-agent-kit, an MIT-licensed open-source Claude Code plugin built on one premise: cheaper models handle I/O and routine implementation, while Opus's context is reserved for judgment and the review gate.
What's inside:
- Seven Claude Code subagents
- Orchestration rules for the main session
- A read-budget hook
- A CLAUDE.md template
Install via /plugin marketplace add FlaviusBurghila/claude-agent-kit; /agent-kit:init writes an "Agent contract" section into CLAUDE.md, and /agent-kit:doctor safely checks the setup. Requires Claude Code 2.1.293+. Fully free, no paid tier.
The author admits no measured cost savings — results depend on how much of your session is I/implementation vs. judgment — and asks the community where they draw the model-split line and whether review is where the stronger model matters most.
More from coding & agent
- Don't run bot teams on scheduled wake-ups: an event-driven agent architecture — b_renat · 2026-10-19
- Claude one-shot rebuilds a structural integrity physics demo from a video — erwincoumans · 2026-10-19
- 300 parallel agents plus a 100ms cheap verifier compress weeks of analyst work into one run — CurieuxExplorer · 2026-10-19
- Local Qwen vision model builds a 3D room game in 26 minutes on a single prompt — Beneficial-Ad-8127 · 2026-10-19
- Building a vast Unreal Engine world from scratch with Opus, says AI game dev is accelerating — sven_ai · 2026-10-19
- DeepSeek V4.1-Flash lands #2 on Terminal-Bench 4.0 at 1/16th the price of Opus 5 — zainhas · 2026-10-19