| name | using-sku |
| description | Use when the agent needs cloud or LLM pricing — AWS, Azure, GCP, or LLM models across 70+ serving providers. Covers point lookups, cross-provider compare, workload cost estimation, and batch multi-op queries. The CLI is offline-first (daily-updated local SQLite shards) and agent-shaped (JSON everywhere, semantic exit codes, field-projection presets). |
Using sku
sku is a pure-Go CLI that answers "what does this cost?" for AWS, Azure, GCP, and LLM serving providers. It is designed for agents: JSON output, semantic exit codes, presets that project to ~200 bytes per record.
Check availability
sku version --pretty
If sku is not installed, direct the user to install.md (brew install sofq/tap/sku, npx @sofq/sku, pipx install sku-cli, scoop install sku, docker pull ghcr.io/sofq/sku).
First-run bootstrap
sku is offline — pricing data ships as daily-updated SQLite shards fetched by sku update.
sku update openrouter aws-ec2
Core verbs
Point lookup
sku aws ec2 price --instance-type m5.large --region us-east-1 --preset agent
sku azure vm price --arm-sku-name Standard_D2_v3 --region eastus --os linux
sku gcp gce price --machine-type n1-standard-2 --region us-east1
sku llm price --model anthropic/claude-opus-4.6 --serving-provider aws-bedrock
Cross-provider compare
sku compare --kind compute.vm --vcpu 4 --memory 16 --regions us-east --limit 5 --preset compare
sku llm compare --model anthropic/claude-opus-4.6 --preset compare
Search within one shard
sku search --provider aws --service ec2 --min-vcpu 4 --max-price 0.10 --sort price --limit 5
Estimate workload cost
sku estimate --item 'aws/ec2:m5.large:region=us-east-1:count=10:hours=730' --pretty
sku estimate --item 'llm:anthropic/claude-opus-4.6:input=1M:output=500K:serving_provider=aws-bedrock' --pretty
Batch many ops in-process
cat <<'EOF' | sku batch
{"command":"aws ec2 price","args":{"instance_type":"m5.large","region":"us-east-1"}}
{"command":"llm price","args":{"model":"anthropic/claude-opus-4.6"}}
EOF
Agent-shape flags (apply globally)
| Flag | Purpose |
|---|
--preset agent (default) | ~200 B record — best default |
--preset price | price[] only — lowest-token lookups |
--jq '<expr>' | gojq filter on the response |
--fields a,b.c | dot-path projection |
--pretty | indent (interactive use only) |
--include-aggregated | include OpenRouter synthetic rows (normally excluded) |
Exit codes are a contract
| Exit | Meaning |
|---|
0 | ok |
3 | not_found (see error.details.nearest_matches) |
4 | validation (see error.details.reason) |
8 | stale_data (run sku update, or pass --stale-ok) |
Full taxonomy: sku schema --errors.
Discovery
Agents should introspect rather than guess:
sku schema --errors
sku schema --list-commands
sku schema --list-serving-providers
Common agent patterns
"Cheapest cloud VM that fits N cores / M GB"
sku compare --kind compute.vm --vcpu N --memory M --regions us-east --limit 1 \
--preset compare --jq '.[0]'
"Cheapest way to serve a given LLM"
sku llm compare --model "$MODEL" \
--jq '[.[] | select(.resource.attributes.aggregated != true)]
| sort_by(.price[0].amount)[0]'
"Budget check before committing to an architecture"
sku estimate --config <workload.yaml> --pretty
Guardrails
- Do not pass
--auto-fetch inside tight loops. It can cause a CDN fetch mid-query. Prefer an explicit sku update up front.
- Aggregated OpenRouter rows (
resource.attributes.aggregated = true) are not real endpoints; exclude them with --include-aggregated off (the default).
- Sku never reaches provider APIs; do not prompt the user for cloud credentials in order to query pricing.
When to stop and tell the user
sku version fails → tell the user to install sku.sku update fails with exit 4 (shard_too_new / binary_too_old) → tell the user to upgrade sku.sku update fails with exit 7 → CDN upstream issue; retry in a few minutes.- Exit code 8 with
--stale-ok already set → data is stale beyond your configured stale_error_days — tell the user.
