Agent Audit

Cut your AI agent costs by 40-60% without losing quality.

Most people running AI agents are overpaying. Opus on a health check. GPT-4 on a reminder. That's $75/M tokens on a task a $0.25 model handles perfectly.

Agent Audit scans your entire OpenClaw setup and tells you exactly where you're bleeding money — and what to swap.

What It Does

• Scans every cron job, agent, and model assignment in your config
• Classifies tasks by complexity — simple tasks get cheap models, complex ones stay protected
• Recommends specific swaps with estimated monthly savings and risk levels
• Works with any provider — Claude, GPT, Gemini, Grok
• Never downgrades coding, security, or trading tasks

Install

npx clawhub install agent-audit

Usage

Once installed, just tell your agent:

"audit my agents" "where am I wasting tokens" "optimize my AI costs"

Or run the script directly:

python3 scripts/audit.py                    # Full markdown report
python3 scripts/audit.py --format summary   # Quick savings number
python3 scripts/audit.py --output report.md # Save to file

Example Output

💰 SAVE $12.40/mo: Swap "Health Check" from opus → haiku
   Current: $15/M input, $75/M output
   Suggested: $0.25/M input, $1.25/M output
   Risk: LOW | Confidence: HIGH

💰 SAVE $8.20/mo: Swap "Weather Check" from sonnet → haiku
   Current: $3/M input, $15/M output
   Suggested: $0.25/M input, $1.25/M output
   Risk: LOW | Confidence: HIGH

Supported Providers

Provider	Models
Anthropic	Opus, Sonnet, Haiku
OpenAI	GPT-4.5, GPT-4o, GPT-4o-mini, o1, o3-mini
Google	Gemini 2.5 Pro, 2.0 Flash, Flash-Lite
xAI	Grok-3, Grok-3-mini, Grok-4

Safety Rails

The audit never recommends downgrading:

• Coding and development tasks
• Security reviews and audits
• Trading and financial tasks
• Tasks that previously failed on weaker models
• Your main interactive session model

How It Works

1. Discovery — reads your OpenClaw config, maps every agent and cron job to its model and pricing
2. Analysis — pulls run history, calculates per-job token usage, runtime, and success rate
3. Classification — categorizes each task as simple, medium, or complex based on name patterns, output length, and runtime
4. Recommendations — flags every mismatch where you're paying more than you need to, with confidence scores and risk levels

Requirements

• OpenClaw installed and configured
• Python 3.8+

License

MIT — free and open source.