spotinfo

Pick a Spot machine on AWS, GCP or Azure from one command line. No credentials. Every price ships inside the binary, so --offline answers in about a tenth of a second.

The problem

Spot capacity is far cheaper than On-Demand — between 40% and 85% off in the snapshots that ship with this tool. Finding the right machine does not scale by hand:

• Each cloud publishes prices in a different place, in a different shape, on a different schedule. AWS has a JSON feed and an interruption Advisor. GCP has server-rendered HTML pages. Azure has a REST API and a separate documentation site for vCPU and memory.
• The consoles rank by price. They do not rank by "cheapest machine with at least 4 vCPU and 16 GiB that my workload can survive losing".
• A price page has no memory of yesterday, so a script that reads one is a scraper you now own.

The result is that most teams pick one instance family, hard-code it, and stop looking.

What spotinfo does

One command ranks real machines against a real requirement, and says where the number came from.

$ spotinfo recommend --cloud azure --architecture arm64 --min-vcpu 4 --min-memory-gib 16
RANK  CLOUD  REGION        MACHINE            ARCHITECTURE  vCPU  MEMORY GiB  USD/HOUR    SAVINGS  RISK          WHY
   1  azure  centralindia  Standard_D4ps_v6   arm64            4        16.0  0.017076        81%  unavailable   ARCHITECTURE_MATCH,COST_POLICY,KNOWN_POSITIVE_PRICE,RESOURCE_MINIMUMS_MET
   2  azure  centralindia  Standard_D4ps_v5   arm64            4        16.0  0.018665        81%  unavailable   ARCHITECTURE_MATCH,COST_POLICY,KNOWN_POSITIVE_PRICE,RESOURCE_MINIMUMS_MET
   3  azure  centralindia  Standard_D4pds_v5  arm64            4        16.0  0.022361        81%  unavailable   ARCHITECTURE_MATCH,COST_POLICY,KNOWN_POSITIVE_PRICE,RESOURCE_MINIMUMS_MET

Four properties make that output worth trusting:

It works with no credentials. Price snapshots ship inside the binary, and a weekly job refreshes them through a reviewed pull request. AWS and Azure additionally read live feeds when they can reach them — AWS on every run, Azure when you name one or two regions — and fall back to the snapshot when they cannot. --offline skips every price and risk request; only --with-score still reaches a cloud, because no snapshot carries a placement figure.

It says what it does not know. GCP and Azure publish no redistributable interruption data, so every candidate reports RISK: unavailable. It is never a zero and never a low bucket. A cloud that measures nothing must not outrank a cloud that measures honestly. On GCP, --live-risk fetches a per-project preemption rate for the ranked page — Google measures it differently from AWS, so it is shown and never filtered on.

It refuses questions it cannot answer, and says why. --workload web caps interruption frequency at an AWS Spot Advisor bucket boundary. Ask for it on a cloud that measures something else and the command stops before it reads a price, naming the vendor limit rather than a feature nobody built:

$ spotinfo recommend --cloud gcp --workload web --architecture x86_64 --min-vcpu 4 --min-memory-gib 16
spotinfo: gcp: unsupported capability: risk: the web workload caps interruption frequency at 5%, an AWS Spot Advisor bucket boundary, and gcp publishes no figure measured that way; workload cost applies no ceiling and answers on every cloud

Every answer carries its source. The JSON report names each source URL and the SHA-256 of the document that was read, so a reader can fetch it again and compare.

Install

# macOS with Homebrew
brew install alexei-led/tap/spotinfo

# Linux and Windows
curl -L https://github.com/alexei-led/spotinfo/releases/latest/download/spotinfo_linux_amd64.tar.gz | tar xz

# Docker
docker pull ghcr.io/alexei-led/spotinfo:latest

macOS, Linux and Windows, on AMD64 and ARM64. Full instructions: Installation.

Use it

# Cheapest Arm machine with 4 vCPU and 16 GiB, on each cloud
spotinfo recommend --cloud aws   --architecture arm64 --min-vcpu 4 --min-memory-gib 16
spotinfo recommend --cloud gcp   --architecture arm64 --min-vcpu 4 --min-memory-gib 16
spotinfo recommend --cloud azure --architecture arm64 --min-vcpu 4 --min-memory-gib 16

# A web tier that must survive interruption: AWS only, interruption capped at 5%
spotinfo recommend --architecture x86_64 --min-vcpu 2 --min-memory-gib 8 --workload web

# The five cheapest AWS regions for a machine. Every region is the default
spotinfo recommend --architecture x86_64 --min-vcpu 4 --min-memory-gib 16 --top 5

# A versioned JSON report for a pipeline
spotinfo recommend --cloud azure --architecture arm64 --min-vcpu 4 --min-memory-gib 16 \
  --output json

# Browse what a cloud publishes, with prices and a risk column
spotinfo list --machine "m5\." --region us-east-1

# Add AWS placement scores. This one needs AWS credentials
spotinfo list --machine "m5\." --region us-east-1 --with-score

Two commands, on purpose, and both answer on all three clouds. spotinfo list requires nothing and answers "what is there". spotinfo recommend requires an architecture and a size floor and answers "what should I run". They share a vocabulary, not a purpose. See Quick start and the Usage guide.

Cloud coverage

Region lists, machine counts and the reasoning behind each limit: Cloud coverage.

MCP server

spotinfo is a Model Context Protocol server, so an assistant can ask these questions directly.

{
  "mcpServers": {
    "spotinfo": { "command": "spotinfo", "args": ["--mcp"] }
  }
}

Ask: "Cheapest arm64 Azure Spot VM with 4 vCPUs and 16 GiB", or "Compare m5.large spot prices across US regions". Three tools answer: list_spot_machines, recommend_spot_machines and list_cloud_regions.

Setup: MCP server and Claude Desktop.

Documentation

AWS credentials

Credentials are optional. Without them, spotinfo answers from the AWS feeds and the embedded snapshot. With them, two more features work:

Credentials load through the standard AWS SDK chain. The chain is probed once per run, so a machine without credentials skips both calls instead of waiting for each to time out.

GCP credentials are optional too. Application Default Credentials and a project — from --gcp-project or GOOGLE_CLOUD_PROJECT — are needed for --live-risk and for --with-score, and a Cloud Billing Catalog API key in --gcp-billing-key prices GCP regions beyond the committed snapshot. Azure needs no credentials at all. See Cloud coverage.

Development

Go 1.26+, make, golangci-lint.

make build            # hermetic: embeds the committed data, downloads nothing
make test             # unit and end-to-end tests, no credentials, no network
make lint
make verify-data      # manifests, source contracts, parser contracts, coverage floors

Contributions are welcome. Read CLAUDE.md for the repository rules. Make sure that every test passes before you open a pull request.

License

Apache 2.0. See LICENSE.