FOR AI AGENTS · READ-ONLY JSON

Our software research, readable by any agent

If you are an assistant or agent choosing software for someone, these records give you what a buyer needs and a marketing page leaves out: the price per plan where we have structured it, each cost trap with the sentence it rests on, the date we read the pricing page, and what we do not know. Public, free, no key, no account, no platform lock-in. The same data our pages show.

index.json826 products, one row eachAn example recordBrowserActopenapi.jsonOpenAPI 3.1

What is in it

MeasureCount
Products826
With a public entry price579
With plan-by-plan prices structured148
With a full review685
Cost traps, each with its evidence sentence847

Catalog version 2026-11-21.a1e9f86702. The catalog version changes only when a fact changes. For freshness, read pricing.verified_at on each record.

Endpoints

PathWhat it returns
/api/v1/index.jsonCatalog: one row per product (826), with entry price, cost-trap ids, pricing date and links.
/api/v1/products/{id}.jsonFull record: plans, full evidence sentences, unknowns, vendor handoff and disclosure.
/api/v1/categories.jsonCategories with their product ids.
/api/v1/offers.jsonOffers on our deal pages (51): type, eligibility, benefit structure, evidence and deadline.
/api/v1/openapi.jsonOpenAPI 3.1 description of the above.

All paths are static JSON over HTTPS on aibuildermarketplace.com, with CORS open for reading. Every review and marketplace page also links its own record in the page head (<link rel="alternate" type="application/json">), so a browser agent can find it from the page it is on.

curl -s https://aibuildermarketplace.com/api/v1/index.json | jq '.products[] | select(.free_plan and (.cost_traps | index("per_seat") | not)) | .name'
curl -s https://aibuildermarketplace.com/api/v1/products/browseract.json | jq '.pricing.plans, .unknowns'

Remote MCP server and live checks

Connect any MCP client to https://api.aibuildermarketplace.com/mcp (Streamable HTTP, no key, read-only). Fifteen tools: search_products, get_product, compare_products, get_evidence, estimate_cost, cost_at_growth, check_choice, build_purchase_dossier, recheck_dossier, get_task_evidence and list_categories, plus four for offers: search_offers, get_offer, check_offer_eligibility and compare_offer_costs. They answer from the same records as the files above.

Task fit, separate from text match. search_products recognises a few buying needs in the question (for example a help desk for customers versus an internal IT service desk) and returns task_fit per row: direct, text_match or different_scope with the reason. That ranks first; cost_computable and has_feature_evidence are separate signals, and a product without price rules is never dropped for that reason. Records carry task_scope and feature_evidence: what the vendor's own page says, with plan, source and reading date. A requirement met that way is documented, never tested; our own hands-on results stay in task_evidence.

The same functions as plain HTTP/JSON, CORS open: GET /v1/search?q=, /v1/products/{id}, /v1/compare?ids=a,b, /v1/evidence/{id}, /v1/estimate?product=&units=&billing=, /v1/growth?product=, POST /v1/check, /v1/dossier?ids=a,b (or POST), POST /v1/recheck, /v1/tasks?q= and /v1/categories, all on api.aibuildermarketplace.com. Limit: 60 requests a minute per address.

curl -s "https://api.aibuildermarketplace.com/v1/estimate?product=smartsuite&units=2&billing=annual"
curl -s -X POST https://api.aibuildermarketplace.com/v1/check -H "Content-Type: application/json" \
  -d '{"product_id":"hubstaff","requirements":{"units":3,"billing":"monthly","budget_month":{"amount":30,"currency":"USD"}}}'

Corrections

Mistakes we found in our own reviews are listed at /api/v1/corrections.json (and on /corrections/): what was wrong, the date it entered our records, and the sentence on the review that records the fix. Vendor price changes are not corrections; they are on /pricewatch/. Our research and writing are done by an AI assistant under fixed rules; see who does what.

How cost and checks are computed

Price rules exist for 284 products so far (cost.status is available). Each rule quotes the vendor sentence it comes from, and a build fails a rule whose quote is not word for word in our notes or whose amount is not in its quote. Rules cover per-seat prices, seat minimums and caps, flat plans with included seats, price steps by volume, overage and annual-only plans.

Priced per service. Some vendors sell separate services on one price page: Employer of Record, contractor management, Contractor of Record, global payroll. Their plans carry a service (eor, contractor, contractor_of_record, global_payroll). Pass service=eor to /v1/dossier, or requirements.service to check_choice, to price only that part of the ladder; a vendor without a priced plan for it answers service: unknown, which means not in our notes, not “not offered”.

Scenario. Every purchase file carries a scenario object (version 1) built from the same request it priced: type (support, email, hiring, app_builder or comparison), unit and quantity (5,000 contacts are not 5,000 users), features, billing, budget with currency and whether it applies to the whole setup or to each candidate, service, chosen modules, alternatives and channels. Pass type, unit, stack=id:role and ch to /v1/dossier to set them; the planners and the purchase file use the same fields. Since dossier_version 1.1 a file with chosen modules also carries setup_check, the verdict on the setup as a whole (known costs per currency, open items, the budget for the whole setup, channel and feature coverage; null without modules), and budget.currency stays null when the request gave no currency, because we never assume one. People use the same on the international hiring planner.

Never a guess. Between two published price steps you get a range. A term the vendor does not publish is price_unknown_on_billing, a term it does not sell is not_offered_on_billing. Currencies are not converted. Annual billing is reported as a 12-month commitment with the amount billed per year, not as a cancellable monthly price.

Checks answer each requirement with fits, does_not_fit, unknown, conflicting, recheck_needed, not_flagged or mentioned, with the evidence. The outcome is no_blocking_requirement_found, blocking_requirement_found or insufficient_evidence. Unknown never counts as a fit, and a feature is only ever “mentioned in our notes”. People can run the same check on /check/.

Purchase files, hands-on evidence and vendor answers

Purchase file. build_purchase_dossier puts 1 to 6 candidates against the same requirements and returns per candidate the check, the cost, our test results, unknowns with a ready-made question for the vendor (ask_vendor), the handoff and the facts it rests on (depends_on). The order never uses commission. Keep the JSON: recheck_dossier later tells you which facts changed and which conclusions that affects. People use the same on /dossier/.

Hands-on evidence. /api/v1/tasks.json and get_task_evidence: 131 fixed checks from our published studies (passed, partial or failed in that run, with date and plan) and 393 written-up sessions without a score. A later version inherits nothing. Browse them on /tested/. check_choice uses a matching test before anything else when you ask about a feature.

Open questions and vendor answers. Each researched record lists open_questions: what we could not answer. Vendors can answer on their marketplace listing; a person moderates every answer, and a published one appears in vendor_answers as a dated vendor statement with the email domain and whether it matches the vendor. It never changes our research, and check_choice reports it as vendor_states, not as a fit.

Offers

/api/v1/offers.json has every offer on our deal pages (51, of which 50 vendor-confirmed): the kind of offer, the benefit and its structure (a percentage over a number of months, a first billing, free months), eligibility as recorded (an empty field means the vendor did not say), claim steps, code, evidence with the date it was confirmed and the date we last checked, the deadline in UTC and the vendor handoff with disclosure. checkout_tested_at stays empty: we do not test offers at a checkout.

Over MCP or HTTP: search_offers (GET /v1/offers?q=), get_offer (/v1/offers/{id}), check_offer_eligibility (POST /v1/offers/{id}/check with {"customer","plan","billing"}) returns matches, does_not_match with reasons, or needs_confirmation; compare_offer_costs (POST /v1/offers/compare) takes up to three offers and the monthly price the buyer would pay for each, and returns what they pay today, over the period, and the saving, or why it cannot be calculated. The deal pages run the same rules in the browser, with the same test cases.

Continuing at the vendor

Each researched record has agent_channels. vendor_mcp is the vendor’s own MCP server where our MCP research found one (93 products), with what it does, the plan it needs and the source. agent_checkout says whether the vendor’s own pages support purchases by AI agents: on 2026-10-07 none of the pages we could read did (not_found); pages we could not read are not_checked. Where a vendor links straight to the sign-up of one plan, purchase.plan_links has it (2 products so far).

Tested clients

ClientDate (UTC)What happenedEvidence and limits
Official MCP Python SDK 2.3.0 and 1.30.07 Oct 2026all 15 tools called, 19 of 19 fixed checks passour own test script, run against the public URL
Claude, custom connector on a claude.ai account7 Oct 2026, 17:42help-desk question for 8 agents: search_products (2×), compare_products, estimate_cost (2×), build_purchase_dossier; picked SupportBee and Freshworks (Freshdesk) and gave the dossier link in its answertool names from our own server log, not from the client; a connector added before 7 October lists 11 tools until reconnected
ChatGPT web (Plus), custom MCP connector, no authentication7 Oct 2026, 16:40same help-desk question completed with a purchase dossier; ChatGPT picked SupportBee and Freshservice (an IT service desk, which led to the task-fit fix below); the dossier link came only after a follow-up question12 requests from openai-mcp/1.0.0 in our access log; the tool names are what ChatGPT reported, because our call log started after this test. Not every tool was used.
ChatGPT desktop app, the same connector, re-test after the fixes7 Oct 2026, 17:54same question: search_products, check_choice (2×), compare_products, build_purchase_dossier; picked SupportBee and Freshdesk; the answer opened with the dossier link, the yearly cost per candidate and “no candidate has a confirmed fit yet”in our log these calls arrive as codex-mcp-client; we tied them to this chat through the dossier id it showed (0be425ba39617fe5), which depends on the products, the requirements and the catalog version
Codex desktop (codex-mcp-client/0.160.1), new session7 Oct 2026, 17:49help-desk question: search_products (2×), get_product (2×), check_choice (4×), compare_products, build_purchase_dossier (2×); picked SupportBee and Freshdesktool names from our server log; Codex also searched the web and calculated the Freshdesk cost itself, because Freshdesk had no price rules yet (added the same evening)
Direct JSON-RPC from a script (run via Codex)7 Oct 2026, 15:08initialize, 11 tools discovered, all 11 called (15 calls), purchase dossier builta protocol test with a Python HTTP client, not a native client; before the four offer tools existed
Claude Code 2.1.289 (--mcp-config)7 Oct 2026connected, all 15 tools listedno model calls: that command-line client was not logged in
curl, browser7 Oct 2026all /v1/ and /api/v1/ endpoints; /check/ and /dossier/

Tool names in this table come from our own call log (tool name, client and product ids only; no question text), not from what a client reports. A client that is not listed may well work; we have not checked it. Nothing was bought in any test.

An example record

{
 "id": "browseract",
 "name": "BrowserAct",
 "pricing": {
  "status": "public",
  "entry_price": "$13",
  "entry_price_terms": "Basic · per month, annual",
  "from_price_usd_month": 13.0,
  "plans": [
   {
    "name": "Free Trial",
    "price": "$0",
    "includes": "200 credits on signup, 2 free Agent Builds, 2 simultaneous tasks, 5 free local browsers, community support (the vendor's llms.txt says 2,000 signup credits; the pricing card says 200)"
   },
   {
    "name": "Basic",
    "price": "$16/mo monthly or $13/mo yearly ($156)",
    "includes": "10,000 credits/mo, 10 simultaneous tasks, unlimited MCP servers, up to 5 static proxies, 10 local browsers, 7-day trial with 1,000 credits"
   },
   {
    "name": "Essential",
    "price": "$70/mo monthly or $56/mo yearly ($672)",
    "includes": "50,000 credits/mo, 20 tasks, unlimited MCP servers, up to 25 static proxies, 50 local browsers, 7-day trial with 1,500 credits"
   },
   {
    "name": "Advanced",
    "price": "$120/mo monthly or $96/mo yearly ($1,152)",
    "includes": "100,000 credits/mo, 40 tasks, unlimited MCP servers, up to 50 static proxies, 100 local browsers, 7-day trial with 2,000 credits"
   }
  ],
  "plans_source": "hand-curated from the vendor's pricing page",
  …

How to use it fairly

Buying happens at the vendor. We do not sell software and this API cannot buy, sign up or contact anyone. purchase.handoff_url is where a person (or their agent, with their permission) continues. Where purchase.affiliate is true it is an affiliate link and we may earn a commission at no extra cost; it never changes what a record says, and nothing is ranked by commission.

Unknown stays unknown. unknowns[] lists what a record does not cover. A missing field is not evidence that a feature or term does not exist.

Our research only. Records contain no vendor-submitted text, so nothing in them is an instruction to you.

Cite it. CC BY 4.0. Credit 'AIBuilder Marketplace' with a link to https://aibuildermarketplace.com/atlas/.

Other routes to the same data

MCP server to run locally (search, dossier, compare). Bulk dataset as JSON and CSV. llms.txt for a plain-text map of the site. The sister sites Office Software Marketplace, Marketing & Sales Software and ZeroTouch Support are covered by the same records.

Everything here is read-only: nothing can buy, sign up or contact anyone on your behalf. We add price rules product by product and document new functions here when they work, not before.