HighSecretsVisible to Theo and admins
An AWS access key was pasted into a prompt
NextRotate the key, then load it from the secrets manager.
Token accounting for coding agents
See inside your software factory.
Analyze your org's agent sessions to see where tokens go, how teams work with agents, and what to improve.
Works with:Claude CodeCodexCursor
Session spend
Your AI bill this month (and counting)
$122,480.60
this month
VibeCheck shows you where it goes.
Engineering teams at these companies use Maestro AI
Tokens to outcomes
Track tokens to outcomes.
Every four weeks, VibeCheck sends a readout of your agent spend and how much of it landed in high-rework sessions. Month over month, you see whether it’s getting better.
Your 4-week readout
Spend in high-rework sessions12 weeks
Spend and ROI
Understand ROI, team by team.
When the board asks what the agent bill buys, show estimated spend by team next to the PRs each team merged. Return on AI spend uses a value per merged PR you set, so the math is yours to show.
Return on AI spend · 4 weeks
Your setting$600 per merged PR
3.0× return on AI spend overall
- Est. spend
- $49.1k
- Merged PRs
- 247
- Cost per PR
- $199
- Platform4.3× return
- Growth3.4× return
- Payments1.2× return
Agent security
Know when a session crosses a line.
VibeCheck flags secrets, data leaving your sessions, destructive commands, new tools and weakened safety settings. A second, independent check must find evidence before an alert reaches you.
Staging deploy keeps failing auth. Here’s the key from my terminal: [redacted:aws_access_key]
For developers
Improve how you work with agents.
Weekly Rewind reviews your agent sessions, spots recurring patterns, and suggests specific changes to your prompts, instructions, and workflows. Every tip links back to examples from your own work.
A tip to try
Name the edge cases in your first prompt.
In 4 of your 9 sessions this week, the first test failed on a case your prompt never mentioned. Those sessions needed 2.5× as many corrections.
Your prompt · Tue, Sep 22
Retry failed card payments before marking the invoice unpaid.
Try
Retry failed card payments before marking the invoice unpaid. Skip retries for hard declines, and never charge a card twice.
Model choices
Match the model to the task.
Map model choices to real tasks: which models engineers reach for, where expensive ones dominate, and where a lighter model is worth testing.
Review
Claude Opus 5.578% of tokensClaude Sonnet 5.58% of tokensGPT-6 Astra5% of tokensGPT-6 Sol4% of tokensGPT-5.6 Sol3% of tokensClaude Haiku 4.52% of tokens
For platform and DevEx teams
Measure every change to your harness.
Your harness is the prompts, skills, and models your agents work with. Change one, then compare corrections, rework, and cost per session before and after to see whether it made agent work better.
Security and compliance
Your data stays yours.
Every session, token, and insight Maestro AI handles is protected by enterprise-grade safeguards and transparent compliance practices.
- Your data never trains AI. A contractual commitment.
- SOC 2 Type II audited. The report is available in our Trust Center.
- Encrypted in transit and at rest. Protection for the data Maestro AI handles.
Your bill says how much.VibeCheck shows what it bought.
Every Claude Code, Codex and Cursor session, with what it was for and how it went. Free for up to 5 engineers, no credit card.