ARC Open

ARC (arcreport.ai) is a free, public database and benchmark for agentic commerce: it measures whether AI shopping agents can actually find, cart and check out at real online stores, across protocols (UCP, ACP, WebMCP, MCP and agent payment programmes) and platforms. It is neutral: not tied to any model, protocol, platform or payment company.

This repository holds the reusable parts of ARC, free for anyone to use, copy and build on:

Folder	What it is	Licence
packages/arc-check	arcreport-check: a zero-dependency CLI, library and GitHub Action that checks a store's agent-commerce readiness (UCP profile, WebMCP, catalog, checkout link, MCP, robots, product data, policies)	MIT
packages/ucp-validator	The UCP business-profile validator: published JSON Schema + spec rules + clearly labelled interop hints	MIT (vendored UCP schemas Apache-2.0)
benchmark/	ARC-100 v1, a fixed set of 100 real stores for comparing shopping agents, the selection code that reproduces it, and the leaderboard submission validator	Data CC-BY-4.0, code MIT
docs/	Methodology and scoring, leaderboard submission spec, dataset schema/card, OpenAPI for ARC's public API, MCP server entry	CC-BY-4.0
scripts/publish_hf.py	Mirrors the weekly open dataset to Hugging Face (--dry-run builds files only; upload reads HF_TOKEN)	MIT

Quickstart

Check any store (Node 18+, no account, no key):

npx arcreport-check yourstore.com
npx arcreport-check yourstore.com --require checkout   # can agents reach checkout?
npx arcreport-check a.com b.com --require all --json

No npm? Same tool as one file: curl -fsSLO https://www.arcreport.ai/arc-check.mjs && node arc-check.mjs yourstore.com

In CI (fails the job if your UCP profile or WebMCP breaks):

- uses: abab7575/arc-open/packages/arc-check@v0.1.0
  with:
    host: yourstore.com
    require: ucp,webmcp

Benchmark your shopping agent: run it on the 100 stores in benchmark/arc-100/v1/arc-100.csv, stop before payment, and submit using the spec.

The data

• Open dataset (weekly snapshots, CSV and Parquet, CC-BY-4.0): https://www.arcreport.ai/data. Field reference: docs/dataset-card.md.
• On Hugging Face: ArcReport/arc-agent-commerce, refreshed weekly: load_dataset("ArcReport/arc-agent-commerce", "rails_checks")
• Developer hub (free API, MCP server, UCP validator): https://www.arcreport.ai/developers
• Method: docs/methodology.md and https://www.arcreport.ai/methodology
• Leaderboard: https://www.arcreport.ai/leaderboard

ARC labels every result Measured, Inferred or Not measured, and never claims more than the evidence shows. ARC never places real orders: every test stops before payment.

Develop

npm ci
npm test         # CLI, UCP validator and ARC-100 reproducibility tests
npm run typecheck

Cite

See CITATION.cff. Short form: ARC (2026). ARC Agent Commerce Index. https://www.arcreport.ai/data

Contributing and contact

Issues and pull requests are welcome: see CONTRIBUTING.md. Store owners can request a correction or opt-out at https://www.arcreport.ai/contact.

Licence

Code: MIT. Data and docs: CC-BY-4.0, attribution "ARC (arcreport.ai)".