bedrock-quota

A CLI-based dashboard to see your AWS Bedrock model quotas and usage.

The AWS console shows quota limits on one page and CloudWatch usage on another — nothing joins them per model. bedrock-quota pulls both and shows headroom per model, per region, so you know where you stand before you scale. Read-only, no dashboard to maintain.

Features

• Know your headroom — RPM, TPM, TPD quotas next to last-day usage, per model.
• Spot throttle risk early — ⚠ flagged when 7-day P90 TPM crosses 80% of quota.
• Every variant in one table — on-demand, cross-region, and global cross-region inference profiles.
• Drill into a model — per-period stats, stacked token charts, hourly heatmap, per-inference-profile breakdown.
• Multi-region view — each region loads independently, shown as a separate section.
• Keyboard-first — works over SSH, 5-minute metric cache.

Install

uvx bedrock-quota                # try without installing (recommended)
uv tool install bedrock-quota    # or install as a uv tool
pipx install bedrock-quota       # or via pipx
pip install bedrock-quota        # or plain pip

Quickstart

AWS_PROFILE=my-sso-profile AWS_DEFAULT_REGION=us-east-1 bedrock-quota

Credentials are picked up like any other AWS CLI tool — AWS_PROFILE, SSO, environment variables, or an attached instance role.

IAM permissions

Most developers already have enough read-only access through their SSO or dev IAM role and can skip this section. If you need to scope a dedicated policy, this is the minimum:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:ListFoundationModels",
        "bedrock:ListInferenceProfiles",
        "bedrock:ListTagsForResource",
        "cloudwatch:GetMetricData",
        "cloudwatch:ListMetrics",
        "pricing:GetProducts",
        "servicequotas:ListServiceQuotas"
      ],
      "Resource": "*"
    }
  ]
}

Keyboard shortcuts

Model detail

Press Enter on any row to open a per-model detail screen with stats (P50/P90/Avg/Max for TPD, RPM, TPM, latency), token charts, an hourly heatmap, and a per-inference-profile breakdown — for time periods Last 24h, Today, 7d, 14d, 30d, This Month, Last Month.

Troubleshooting

• No usage data — CloudWatch only records metrics for models you've actually invoked. Check that AWS_DEFAULT_REGION matches the region you're calling Bedrock in.
• Slow first load — the first run queries CloudWatch for every model across every time period; subsequent loads hit the 5-minute cache. Press r to force-refresh.
• Credentials error — run aws configure, aws sso login, or set AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY / AWS_DEFAULT_REGION.

Implementation notes

Usage data is queried from CloudWatch (AWS/Bedrock namespace, ModelId dimension): InputTokenCount + OutputTokenCount → TPM/TPD, Invocations → RPM, InvocationLatency → latency, InvocationThrottles / InvocationClientErrors / InvocationServerErrors → error counts. Quotas come from AWS Service Quotas. Model IDs are discovered from ListFoundationModels, ListInferenceProfiles, and observed CloudWatch dimensions, so cross-region variants like us.anthropic.claude-sonnet-4-6 match their quota automatically. Per-token costs shown in the detail screen are list prices from the AWS Pricing API (pricing:GetProducts); they exclude Savings Plans and batch-inference discounts. If pricing:GetProducts is not granted, cost rows render — and the status bar shows a note.

Development

git clone https://github.com/Kelet-ai/bedrock-quota
cd bedrock-quota
uv sync
uv run bedrock-quota

License

MIT