GitHub - shenli/AgentDog: Real-time monitoring dashboard for local AI agents.

See what your AI agents are actually spending.

A lightweight macOS app that monitors Claude Code, OpenAI Codex, and OpenClaw sessions in real time. Token usage, costs, context window health, tool call patterns — all in one dashboard. No API keys. No agent modifications. Just install and go.

Install

Download the latest release (macOS Apple Silicon) — open the DMG and drag to Applications.

On first launch, macOS may block the app. Go to System Settings > Privacy & Security and click Open Anyway.

The Problem

You're running coding agents for hours, but you can't see what's happening:

• A session burns through your token quota and you only notice when it starts throttling
• The context window fills up silently — your agent breaks mid-task with no warning
• A config change breaks prompt caching and costs spike 5x
• A single git log --stat eats 30% of your context and nobody tells you
• You're running multiple agents across projects with no unified view of total spend

What AgentDog Shows You

• Token and cost breakdown per turn — input, output, cache read/write
• Context overflow warnings before compaction happens
• Cache efficiency tracking with invalidation detection
• Tool cost ranking — which tools consume the most tokens
• Anomaly detection — flags turns that cost 3x+ the running average
• Daily spending trends across all agents and sessions
• Provider status — live Anthropic and OpenAI operational status

How It Works

AgentDog reads the transcript files that agents already write to disk:

Read-only. Runs locally. Never writes to agent directories. Data stored in ~/.agentdog/data.sqlite.

Build from Source

Requires Rust and Node.js 18+.

git clone https://github.com/shenli/AgentDog.git
cd AgentDog
npm install
npm run dev

Contributing

Contributions welcome. Please open an issue before submitting large changes.

License

Apache 2.0