I built an AI coding agent 50% cheaper than Claude Code (same prompts)
Minifies code before sending to LLM, slashing Claude Code costs by 50% in benchmarks.
Kon is a minimal coding agent (and also a highly opinionated one)
270-token system prompt beats Cursor's bloat by forcing strict context limits.
Developers frustrated with bloated AI coding assistants
Cursor · Aider · Open Interpreter
Minifies code before sending to LLM, slashing Claude Code costs by 50% in benchmarks.
Replaces UUIDs with space-separated words to slash token costs in LLM prompts and tool calls.
Like top for AI agents: tracks token costs across 6 coding agents, 100% offline.
Drag‑and‑drop phases and agent cards with a live prompt preview is a smart ergonomics win — you can wire agents manually or rely on smart auto-connections and export a CLAUDE.md or .opaal file. The app shows attention to UX (keyboard shortcuts, 50-level undo, auto-detect of local Claude Code skills), but it's an Electron MVP tightly focused on the Claude ecosystem, so its appeal outside that niche is limited.
OTP supervisor trees mapping to agent workflows is genuinely clever, but early-stage and GitHub Copilot-only.
60 DAUs saved 10.5B tokens — real savings for Claude Code power users.