# Machineble — full text corpus > Machineble sends real AI agents to e-commerce sites, measures where they succeed or fail, and reports agent-readiness. Operated by DigitalPunkers s.r.o., Czech Republic. Marketing site: https://machineble.com. Application: https://app.machineble.com. Contact: hello@machineble.com. ## What Machineble is AI agents already shop. They read your site the way a machine does: structured data, feeds, protocol endpoints, robots rules — not your hero image. When they cannot parse a price, resolve availability, or complete a cart, the sale silently goes to a competitor that is machine-readable. Machineble measures both halves of that problem: 1. Protocol readiness — a deterministic scan of the machine surface of your domain. 2. Agent runs — real LLM agents (OpenAI and Anthropic models) attempting real shopping tasks on your live store, scored by task success rate (TSR). The output is a 0–100 agent-readiness score, a per-check breakdown across four quality dimensions (presence, validity, completeness, freshness), a list of findings with severity, business impact and a concrete fix, plus an AI-written executive summary and repair plan. ## Protocol readiness checks Six checks, weighted to 100 points: - schema.org Product structured data (weight 20). Does a product page emit valid JSON-LD with name, image, offers.price, offers.priceCurrency and offers.availability? Recommended: brand, sku, gtin, aggregateRating, review, description. Without it, agents cannot extract product identity or price and fall back to scraping. - llms.txt (weight 15). A curated map of canonical pages per the llmstxt.org spec: H1 title, blockquote summary, `## Docs` section with real links, plus an optional `/llms-full.txt` companion. Without it, LLM crawlers index navigation and cart pages instead of your catalogue. - MCP endpoint (weight 15). A Model Context Protocol server at `/mcp` or `/.well-known/mcp` returning `protocolVersion`, `serverInfo` and a populated `tools` array with descriptions and declared `inputSchema.required`. Agents pick tools by description; a tool-less or thin MCP is functionally invisible. - Universal Commerce Protocol, UCP (weight 20). A discovery document at `/.well-known/ucp.json` declaring version, capabilities and a feed URL, plus an item feed carrying id, title, url, price, availability and image, ideally with gtin, brand, variants, shipping and currency. - Agentic Commerce Protocol, ACP (weight 20). A discovery document at `/.well-known/agentic-commerce.json` declaring `agentic_commerce_version`, merchant and capabilities, plus a feed whose items carry id, link, title, price and inventory_quantity, ideally with enable_search, merchant, attributes, image_link and brand. A stock Shopify `/products.json` alone does not qualify. - Agent access, robots.txt (weight 10). Are GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, anthropic-ai, PerplexityBot, Google-Extended, CCBot, Applebot-Extended, Bytespider, Meta-ExternalAgent and cohere-ai allowed, and is a sitemap declared? Blocking an agent means every surface it powers cannot recommend or buy from you. ## Agent runs and TSR Agent runs execute scripted commerce tasks — find a product, compare variants, add to cart, reach checkout — using real models. Task success rate (TSR) is the share of attempted tasks completed. Failures are categorised (discovery, parsing, navigation, checkout) and paired with the protocol findings that most likely caused them, plus a revenue-at-risk estimate. Currently available models: GPT-5, GPT-4.1 mini and GPT-4o mini from OpenAI; Claude Sonnet 4.5 from Anthropic. More models are in progress. ## How an audit works 1. Discovery — the scanner probes the machine surface of the domain over HTTPS with a declared user agent. 2. Scoring — every check returns presence, validity, completeness and freshness scores, weighted into a 0–100 total. 3. Analysis — an AI analyst turns raw findings into an executive summary and a prioritised repair plan written for a technical e-commerce lead. 4. Report — results are available online and as PDF, and public audits are indexed at https://machineble.com/audits. A free audit takes about 30 seconds and needs no account. Free audits are rate limited per IP address to keep the service available. ## The readable → usable → buyable progression - Readable: an agent can parse what you sell. Structured data and llms.txt. - Usable: an agent can act on your site. MCP tools, clean navigation, deterministic state. This is where most stores fail and where Machineble focuses. - Buyable: an agent can complete a purchase. UCP and ACP surfaces, programmatic cart and checkout. ## Pricing - Free — $0. 1 domain, 1 protocol readiness audit, 3 agent runs, basic online report, GPT-4o mini. No card required. - Pro — $49 per month, or $290 per year (about $24 per month, billed annually). 3 domains, unlimited protocol audits, 50 agent runs per month, AI failure analysis and fix plan, TSR and revenue-at-risk metrics, reports online and PDF, 3 team seats, models GPT-5, GPT-4.1 mini and Claude Sonnet 4.5. - Agency — $149 per month, or $890 per year (about $74 per month, billed annually). 100 domains, unlimited protocol audits, 500 agent runs per month, advanced AI analysis and repair plans, TSR, revenue-at-risk and trend deltas, white-label reports online and PDF, unlimited team seats, API access, all models with priority queue. All plans are self-serve at https://app.machineble.com. Prices are in USD, excluding VAT where applicable. ## Machine interfaces published by Machineble - https://machineble.com/mcp — MCP server. GET returns the handshake; POST speaks JSON-RPC 2.0 (`initialize`, `tools/list`, `tools/call`). Tools: `run_protocol_audit`, `get_audit`, `list_recent_audits`, `get_pricing`. - https://machineble.com/.well-known/agent-access.json — agent access policy, rate limits and contact. - https://machineble.com/.well-known/ucp.json and https://machineble.com/ucp/feed — UCP discovery and plan feed. - https://machineble.com/.well-known/agentic-commerce.json and https://machineble.com/acp/feed — ACP discovery and plan feed. - https://machineble.com/sitemap.xml — all indexable URLs. - https://machineble.com/llms.txt — curated LLM map. ## Company and legal Machineble is a product of DigitalPunkers s.r.o., Czech Republic. Terms of service: https://machineble.com/terms. Privacy policy: https://machineble.com/privacy. Cookie policy: https://machineble.com/cookies. Audit results submitted through the public form may be published in the public audit index at https://machineble.com/audits.