AgentPay: budget guard for coding agents

A real-time spend brake for AI coding agent API bills

Problem

A developer on V2EX reports spending tens of thousands of RMB on API calls, burning 20B tokens in a month, and splitting work by model (Opus for frontend, GPT for backend, DeepSeek for CI). In the same thread people complain that some models loop on verification and waste tokens. Today you only see the bill afterward; nothing stops a runaway agent while it is burning.

Solution

A thin gateway/SDK: route model API calls through it, set budget and rate caps per project, agent and task, and automatically pause, downgrade to a cheaper model, or ask for human approval when a limit is hit. A live dashboard shows spend per agent.

Users

Indie developers and small teams heavily using Claude Code / Codex-style coding agents

MVP

Drop-in proxy endpoint compatible with major model APIs; daily budget and token caps per project/agent; auto-pause plus notification on breach; live spend dashboard broken down by model; export for finance or retro

Revenue

Free tier limited to one project; monthly Pro tiers by agent count or call volume; optional per-channel API billing

Why now

Coding agents became daily infrastructure so cost shifted from per-call to per-day, yet vendor consoles remain post-hoc; one developer can build a proxy and threshold layer in weeks.