Agent Cost Guard

Monitor and stop AI agent budget blowouts from silent cache misses

Problem

AI coding agents silently re-bill entire conversations at cold-write prices when prompt cache breaks, with zero errors in logs. A V2EX post shares a 4-step method to reduce token burn on long agent tasks, showing deep user sensitivity to runaway spend; separate projects like Replay and Dep-radius are already tackling cache audits and dependency impact analysis, confirming a real engineering pain point.

Solution

A local-first CLI and dashboard that parses agent disk logs and API responses, detects prompt cache misses, and alerts when hit rates drop; it breaks down dollar cost per session and enforces hard budget caps for tools like Claude Code, Cursor, and Aider.

Users

Indie developers, small engineering teams, and AI product builders who run coding agents heavily

MVP

Parse local log files from Claude Code / Cursor / Aider; compute per-session token usage and USD cost; detect and highlight cache-miss timestamps; show cost trends in a terminal or lightweight web UI; allow per-project budget thresholds with alerts

Revenue

Free basic monitoring; paid tier at $9–19/month for team dashboards, budget enforcement, and Slack/Discord alerts

Why now

AI coding agents are exploding in adoption, but cost observability lags far behind; big labs won't build cross-tool, local-first cost diagnostics, leaving a clear gap for indie builders