The Agent Readiness
Census.

Publisher: Agent Outposts · Data: registry_seed.json · Checked: 2026-11-01

How open are America's biggest consumer brands to AI agents acting on a customer's behalf? We audited 43 major brands across 9 categories, scoring five consumer intents for each — browsing, prices, inventory/availability, ordering, and tracking — strictly from sanctioned, publicly documented channels. The short answer: the agent economy currently runs on aggregators, not on brands. Only 3 of 43 brands offer any self-serve API door at all, and exactly one brand in the entire census lets an agent complete a purchase end-to-end.

Universe. The fixed brand list: Pizza (4), Coffee (4), QSR (7), Grocery & warehouse (7), Retail & e-commerce (7), Pharmacy (2), Airlines (4), Hotels & stays (4), Platforms (4) — 43 brands, all completed. (The project brief described "~50 brands"; the fixed list enumerates 43, and all 43 are covered.)

Rubric. Each intent is scored with one door-state:

Checks per brand: (a) developer portal / API docs; (b) agent-native surfaces — an official MCP server, /llms.txt, /.well-known/api-catalog, published OpenAPI; (c) aggregator presence; (d) bot posture from public signals only (the brand's own robots.txt, published automation/API terms); (e) sanctioned channels (affiliate/partner programs).

Evidence protocol. Every door-state carries at least one source URL from an official brand/portal page (or the brand's own robots.txt) and the check date. human_only was recorded only after checking both the developer-portal angle and agent-native surfaces. Third-party sources were used only as corroboration; official pages decided. Research was light-touch (≤ ~8 page fetches per brand, no logins, no form submissions).

Limitations — read these before citing the numbers.

• Inventory is an availability proxy in food categories. No pizza, coffee, or QSR brand exposes a true stock API; AG scores for inventory mean "store/item availability as shown by the aggregator," not a live inventory feed.
• Agent-native checks are partially unverified. A browser rate-limit (HTTP 429) interrupted several robots.txt / llms.txt / api-catalog probe batches — most notably the entire Grocery category. llms_txt is "unverified" for 18 of 43 brands and api_catalog for 19 of 43; confirmed absences are distinguished from un-checked ones in the data, and the headline findings below rely only on confirmed positives.
• Two portal fetches failed outright. developer.kroger.com and parts of Instacart's docs were unreachable from our tooling (network/DNS errors), including in an independent coordinator spot-check of Kroger. Those scores rest on the portals' indexed official documentation plus multiple independent implementations citing the exact official URLs, are disclosed in the data's note fields, and are flagged for re-verification.
• Airlines and hotels are partner ecosystems by design. NDC and distribution APIs are real, sanctioned doors — but they are B2B doors. We score them PG for consumer intents and say so, rather than pretending a GDS contract is an agent door.
• This is a snapshot. Doors are moving (DoorDash's MCP is in private beta; Delta's full NDC launch is planned for end of 2026). Every claim carries its check date for a reason.

The census covers 43 brands × 5 intents = 215 brand-intents. Their distribution:

Openness by intent (share of 43 brands with an open_api door):

Findings:

1. Self-serve doors are nearly nonexistent. Only 3 of 43 brands (7.0%) have any open_api intent: Peet's Coffee, Kroger, and Best Buy. Catalog data (menus, prices, availability) is where the few open doors live — 3 brands each for browse/prices/inventory. The transactional intents collapse: ordering and tracking are open at exactly one brand each (2.3%) — Peet's, both times.
2. The aggregator is the de-facto agent door. 97 of 215 intents (45.1%) are reachable only through a third party — DoorDash, Uber Eats, Grubhub, Instacart, Shipt, Google Shopping, Duffel. 16 brands (37.2%) are aggregator-only across all five intents, including Starbucks, McDonald's, Domino's, and Costco. Brands have, in effect, outsourced their agent channel to platforms they don't control.
3. Nearly half the list has no direct relationship available at all. 20 of 43 brands (46.5%) offer neither an open nor a partner door for any consumer intent. 21 brands (48.8%) have at least one partner-gated door — the partnership BD motion is the only game in town for direct access.
4. Exactly one brand is fully walled by policy: Trader Joe's — human_only across all five intents, per its own official FAQ (no online sales, no delivery, no third-party services, no authorized online sellers). It is also the only brand whose wall is a stated philosophy rather than an absence.
5. Agent-native surfaces are embryonic. Confirmed official MCP servers: 3 of 43 (Peet's via UCP, Instacart, DoorDash — DoorDash's is a private beta). Confirmed /llms.txt: 3 of 43 (Peet's, CVS, Dutch Bros). Confirmed /.well-known/api-catalog: 0 of 43. The infrastructure agents natively speak is essentially unbuilt outside a Shopify-driven retail edge.
6. Ordering is the wall that matters. 9 brands (20.9%) are human-only for ordering and 11 (25.6%) for tracking, including Amazon for both — the post-PA-API Creators API covers catalog data for sales-qualified affiliates, but an agent cannot buy or track anything on Amazon through any sanctioned channel.

Codes: OA open_api · PG partner_gated · AG aggregator_only · HO human_only · UV unverified.

Pizza

Coffee

QSR

Grocery & warehouse

Retail & e-commerce

Pharmacy

Airlines

Hotels & stays

Platforms

1. Peet's Coffee — the census's only 5-for-5. Via Shopify's agent stack, Peet's publishes an official llms.txt ("Agent Instructions") and a live /.well-known/ucp profile advertising an MCP endpoint with catalog search, cart, checkout, and order capabilities — all five intents open_api. Scope caveat, recorded in the data: this covers the retail store (packaged coffee and merch); coffeebar drinks remain app/aggregator territory. Still: a mid-size coffee chain is, today, the most agent-ready consumer brand in America on this list — because its platform made it so.
2. Kroger — the grocer exception. A genuinely public, self-serve developer API (Products, Locations) covering browse, store-specific prices, and fulfillment availability (OA ×3), with a public add-to-cart API and partner-gated full cart/delivery. Tracking is the lone unverified cell in the census (no tracking API appears in public or partner product lists).
3. Best Buy — the quiet incumbent. Self-serve Products/Stores/Buying Options APIs (OA for browse, prices, availability) that have existed for years. Ordering stops at an addToCartUrl handoff to the human site — the classic "read-open, transact-closed" pattern.
4. Instacart — the platform with an MCP server. Partner-gated across the board, but it operates a first-party remote MCP server and a public developer platform; as the aggregator door for Costco, Albertsons, CVS/Walgreens retail and more, it is the single highest-leverage integration in grocery.
5. DoorDash — gated, but building the real thing. Its official MCP (find restaurants → menu → preview order with pricing → submit order → tracking) is the most complete agent ordering flow in the census — currently a private beta for approved testers. When it opens, the Platforms row changes character.
6. Walgreens & Papa John's — honorable mentions. Walgreens runs the broadest legacy developer gallery (store inventory, Rx refill/transfer, photo) — all key-gated partner access. Papa John's is the only pizza chain with a documented partner commerce API (menu/prices/ordering via its Commerce Platform portal).

• Trader Joe's — the only fully human_only brand (5/5), by explicit, published choice. Its wall will not move for an Outpost pitch; it is in the census as the limiting case, not a target.
• Starbucks — arguably the highest-frequency purchase intent in the country, with no developer portal, no MCP, no llms.txt: aggregator-only across all five intents. The demand is enormous and entirely intermediated by delivery platforms.
• The QSR wall is total. All 7 QSR brands are aggregator-only across all five intents — 35 of 35 cells. McDonald's and Chipotle pair this with the category's most explicit anti-automation postures (terms banning spiders/robots; Chipotle additionally sits behind a Kasada bot wall with a published text/data-mining ban). Franchise/internal systems (BK Gateway, Byte by Yum!) are not consumer doors.
• Amazon — the most instructive wall: catalog data survives only as a partner door for Associates accounts with ≥10 qualifying sales in the trailing 30 days (the Creators API, PA-API's successor), while ordering and tracking are simply human-only. The largest commerce surface in the census offers an agent no way to transact.
• Costco & Target — top-tier retail demand, zero consumer API; both reachable only through a single owned/partner pipe (Instacart same-day; Shipt respectively), which means the pipe, not the brand, owns the agent relationship.

• Peet's UCP profile — nobody scouting "coffee chains" expects the category's smallest national player to be the census's only end-to-end agent-transactable brand. Platform leverage (Shopify) beat brand scale.
• Delta vs. United, in robots.txt — Delta's robots.txt carries a header "LLM / AI Crawlers – Explicitly Permitted," allowing GPTBot, ClaudeBot, ChatGPT-User, PerplexityBot and others Allow: /, while United's disallows AI crawlers from its shopping paths. Same industry, opposite stated postures toward agent reading — even though both are identically partner-gated for transacting.
• CVS publishes an llms.txt — a substantive one — while having no MCP and a partner-only API program; and Dutch Bros has one too (core pages only). Pharmacy and a drive-thru coffee chain were not obvious llms.txt adopters.
• Kroger's public API exists at all — a top-5 US grocer running a free self-serve products/prices/availability API is the exception that proves the grocery rule (every other grocer on the list: partner, aggregator, or nothing).
• eBay's robots.txt names the use case — it explicitly bans "buy-for-me" / LLM checkout agents even while eBay operates a full Buy API suite behind partner approval: the demand is visible enough to be worth prohibiting.

Ranked by brand size × completeness of the wall × likely agent demand (an Outpost is a demand-registration door at a brand agents already try and fail to transact with):

1. Amazon — The largest stranded demand pool in commerce: ordering and tracking human-only, catalog gated behind sales-qualified affiliate status. Every agent shopping task ends here.
2. Starbucks — Highest-frequency food/drink intent in the US; zero direct door (AG ×5). Coffee runs are a canonical recurring agent task.
3. McDonald's — Largest QSR system; AG ×5 with explicit anti-bot terms. Sheer task volume makes the demand ledger valuable fastest here.
4. Domino's — Pizza ordering is the textbook agent task and Domino's built its brand on digital ordering (and tracking), yet offers no API — AG ×5 via the very platforms it competes with.
5. Costco — High-basket, high-loyalty warehouse demand; no consumer API of any kind; a single Instacart pipe intermediates everything.
6. Walmart — Largest US retailer/grocer; its real APIs serve sellers and affiliates, while consumer ordering and tracking are human-only — a partner program exists to point an embassy conversation at.
7. Target — Top-tier retail demand, no consumer API (Target Plus is invite-only and seller-side); Shipt-only routing.
8. United Airlines — Flight booking is among the highest-value agent tasks; all doors are NDC/aggregator partner contracts, and its robots.txt actively blocks AI crawlers on shopping paths — the sharpest travel wall.
9. Airbnb — Stays are natively agent-shaped demand (search, compare, book, manage); consumer access is partner-only (its PG cells are host-side), and the brand is famously anti-automation.
10. Chipotle — A digital-first QSR (a large share of sales is already digital) behind an enterprise bot wall with an explicit data-mining ban — proof the wall is deliberate, and therefore a decision that can be revisited with demand data in hand.

Notable exclusion: Trader Joe's is the most walled brand in the census but is deliberately not an Outpost target — its no-ecommerce stance is a stated brand philosophy, so expected conversion is near zero.

1. Aggregators are first-class routes, not fallbacks. With 45.1% of all intents aggregator-only — and 16 brands reachable only that way — the registry's routing graph must treat DoorDash, Uber Eats, Grubhub, Instacart, Shipt, Google Shopping and Duffel as primary doors with their own door-states, not as footnotes on brand records.
2. Read doors and transact doors are different products. Browse/prices/inventory have the only open APIs (Peet's, Kroger, Best Buy); ordering/tracking are open almost nowhere (1 brand). The registry should expose this split explicitly — agents can plan research tasks against open doors today, but transactional routing is a partner-gating problem for the foreseeable future.
3. The registry's freshness discipline matters more than its size. Several scored doors are explicitly in motion (DoorDash MCP private beta, Delta NDC full launch planned end-2026, Kroger's partner cart). Door-states need re-verification cadences, and the 37 unverified agent-native fields (18 llms.txt, 19 api-catalog) should be burned down first — they are cheap fetches that convert "unknown" into routable fact.
4. Outpost targets select themselves from this data. The 20 brands with no direct door include most of the highest-demand names in US commerce. The census gives each Outpost page its founding fact — "0 of 5 intents available to agents through any sanctioned channel, checked 2026-11-01" — which is exactly the kind of stamp-true claim the Outpost model runs on.
5. Watch the platform layer, not just brands. The three confirmed MCP servers belong to a platform (Instacart), a platform (DoorDash), and a platform-powered storefront (Peet's via Shopify/UCP). The fastest way doors open across many brands at once is a platform flipping a switch — the registry should track platform-level agent programs as entities in their own right.

Methodology, rubric, evidence protocol and limitations are in the Methodology section above. Per-brand evidence URLs, notes, bot postures and agent-native details are in the companion dataset, registry_seed.json (43 brands, 215 scored intents, checked 2026-11-01).