What a grounded category assistant actually needs#
"What is happening in my category?" is one question to a brand team and four different procurement problems to IT. It decomposes into measured retail sales (share, velocity, distribution such as ACV and TDP); consumer signal (what people cook, search for, post about and say); reference data (nutrients, ingredients, claims, labelling rules); and events about other companies (recalls, enforcement actions, launches).
Those four have different shapes. The first two are commercial licences. An MCP server is a transport, not a licence — if your syndicated contract does not permit programmatic or agent access, no connector changes that, and if it does, the absence of a connector is an engineering problem with a known cost. Get legal to answer the licence question before anyone scopes an integration.
Grounding means something narrower than "the assistant knows about our category". It means the assistant returns a figure it did not invent, attached to something a human could check. That constrains the tool surface you need: tools that return values and identifiers, not an endpoint that returns prose. A server that answers in paragraphs cannot be audited by the analyst who has to defend the number in a category review.
Before you hold credentials, the only pre-contract evidence of that tool surface is what a server publishes: its tool list, if the endpoint answers an unauthenticated request, and its OAuth scope list. Those two fields do more work in an evaluation than any feature page. Across all 33 servers we profile, 16 published a machine-readable scope list to our probe of their protected-resource metadata. Of those 16, eight enumerate scopes that do not separate reading from writing — in some cases a single scope such as Sanity's global or Notion's default, in others a set like Zapier's openid, profile and email that gates identity rather than capability — so read-only access cannot be granted at all, and approving the connection approves everything the server can do. The other eight do separate the two. For the remaining 17 servers, no scope list was observed, so the question is undocumented rather than answered.
What the catalogue contains today#
Five of the 33 servers sit in our market and consumer intelligence category, and the honest reading is that they split three ways: two reference-data sources, one live retail-price source, and two consumer applications that a category team has no use for. Naming all five as though they served the same reader would be padding.
fmcg.network (v0.6.1) is the closest thing here to an FMCG reference source. Its publisher describes USDA FoodData Central nutrient lookup, FDA recall watch and EU FMCG labelling, served from a single hosted endpoint at mcp.fmcg.network/mcp. Our unauthenticated probe on 19 September 2026 got an HTTP 401 with a complete OAuth chain — protected-resource metadata, authorization-server metadata and dynamic client registration all present. It published no scope list to that probe, so whether read-only access can be granted is not documented. Eight of the 18 criteria we track are documented for it.
MyPlate.food (v1.0.0) covers USDA nutrition calculators and the preserved MyPlate Kitchen recipes, described by the publisher as source-cited and machine-verified. Our probe reached it with no credentials at all on 19 September 2026 (HTTP 200); we did not record a tool list from that connection. It declares two endpoints, streamable HTTP and SSE. For grounding a nutrient or claim statement it is directly usable. For anything about what a category is doing commercially, it is the wrong instrument.
offerhopper.ai (v1.0.3) is the only live retail-price source in the catalogue, and it is Germany-only: grocery and drugstore prices, deals and multi-stop shopping routes. Our unauthenticated probe returned exactly two tools — plan_optimal_shopping_route and swap_route_item. Both are route-planning calls. If your team wants a price lookup it can call directly and log, confirm that one exists before designing a price-tracking workflow on this server.
The two we leave out are Spoonjoy and January AI Nutrition. Spoonjoy's own scope list, read from its protected-resource metadata, says what it is: account:read, account:write, cookbooks:read, kitchen:read, kitchen:write, public:read, recipes:read, shopping_list:read, shopping_list:write. That is a personal recipe kitchen, cleanly built — nine scopes with read and write properly separated — and irrelevant to a category review. January AI serves food photo recognition, nutrition search, meal logging and glucose prediction for health apps; it is a developer API for consumer health products, and seven of 18 criteria are documented.
The genuinely useful find sits in a different category. Food safety and compliance holds four servers, and two of them answer questions about other companies, which almost nothing else in the catalogue does. Argus HQ (v1.0.0) exposed five tools to an unauthenticated probe: search_enforcement, list_recent_actions, get_warning_letter, get_recall_by_lot and get_company_history. RecallRadar (v1.0.0) exposed four: search_recalls, get_recall, list_agencies and diff_since. get_company_history is competitive intelligence by another name, and diff_since is what turns a lookup into a monitoring feed — an assistant can ask what changed since a date instead of re-reading the corpus. The other two in that category are credential-gated, so we enumerated no tools for either: atlasverified.ai, which its publisher describes as organic supply-chain verification covering certification, OFAC and FDA import checks, and DAG Studio, which is causal-inference tooling rather than a data source.
None of these seven servers describes, or exposed to our probe a tool for, a sales figure, a share point, a distribution measure or a velocity number, for any market. That is not a gap in our research. It is what the 33 servers we profile, plus the 20 named vendors we checked for and did not find, actually amount to.
The hole where syndicated data should be#
We publish the list of vendors we checked for and did not find, so the finding can be disproved rather than taken on trust. Under syndicated retail data it names five: Nielsen, NielsenIQ, Circana, SPINS and Numerator. None appears in the registry snapshot of 33,512 servers taken on 19 September 2026. Across four groups — syndicated data, supply-chain planning, retail media and foodservice POS — 20 named vendors are absent.
The practical consequence is blunt. The question a category team asks most often is the one with no off-the-shelf connector, and it will stay that way until a licensor decides machine access is a product it wants to sell rather than a leak it has to police.
Our reading of why, offered as reasoning and not as a vendor statement: syndicated data is sold per category, per market, per period. An MCP endpoint turns entitlement enforcement into a query-time problem for the vendor, on a query stream they do not control and cannot easily predict. That is a harder product decision than shipping an API, and it is the part a connector cannot route around.
Disclosure, because it matters here: MCP Compare is built and owned by Tastewise, which sells food and beverage consumer intelligence. Tastewise documents an MCP publicly at tastewise.io/integrations. It is not in the public MCP Registry. That means it has no registry record, no namespace verification, no probe result from us and no profile in this catalogue — it is not scored, ranked or compared on this page, because we hold none of the evidence we require of the 33 servers that are. Access is arranged by contact rather than self-service, which for an IT evaluator means a commercial conversation precedes the technical one. Treat that as a constraint on your timeline, not as a recommendation from us.
What to do in the meantime is unglamorous and effective. Put three questions to your existing data licensor in writing: does our current licence permit programmatic or agent access to the data; is there an API today and under what entitlement model; is an MCP endpoint on the roadmap and when. Then put the answers in the renewal file. Renewal is where demand is registered, and on the evidence here the demand has not yet been registered anywhere the registry can see it.
Carrying data you already licence#
The architecture that works today is not a market-data connector. It is: land the extracts you already pay for in a store you control, model them, and point an MCP server at that store. The assistant then grounds on your data under your entitlements, and the MCP layer becomes an access-control question rather than a procurement one.
Supabase (v0.13.0) is the most thoroughly evidenced option in the catalogue for that role: 13 of the 18 criteria we track are documented, more than any other server here. It scores 100 on our curation scale, is Apache-2.0 licensed, and ships both a hosted endpoint at mcp.supabase.com/mcp and an npm package (@supabase/mcp-server-supabase) you can run yourself. The 13 scopes we read from its protected-resource metadata separate database:read from database:write and include analytics:read, so an insights assistant can genuinely be granted read-only access — one of the eight servers of 33 where that is verifiably true.
Microsoft Fabric MCP Server (v1.4.0) takes the opposite shape and, for a residency-constrained team, a more interesting one. It declares no hosted endpoint at all — only an npm package (@microsoft/fabric-mcp) and a NuGet package (Microsoft.Fabric.Mcp), MIT-licensed. If your syndicated extracts already land in Fabric, that absence is the feature: there is no vendor MCP endpoint for the query to traverse. Be precise about why, though. That is a property of no remote endpoint existing, not of a documented residency control. Its registry record documents nothing about authentication, and with no endpoint there was nothing for us to probe. Six of the 33 servers publish an installable package; this is one.
The Power BI Modeling MCP Server covers the semantic layer where a great deal of CPG category reporting already lives, and is worth a pilot on that basis alone. It is pre-1.0 (0.5.0-beta.13), self-host only via npm, with no hosted endpoint and no licence we were able to determine. Treat it as a spike, not a platform commitment.
Airtable (v0.1.0) is where many category and innovation teams actually keep tracker, launch and competitor-watch data, whatever the architecture diagram says. It is hosted only, with seven scopes in its protected-resource metadata that separate reading from writing — data.records:read, data.records:write, schema.bases:read, schema.bases:write, data.recordComments:read, data.recordComments:write and workspacesAndBases:read — so a read-only grant is available. Amplitude (v1.0.0) is first-party product analytics rather than category data: its publisher describes it as search, access and insight over your own Amplitude data. It declares two hosted endpoints including an EU one at mcp.eu.amplitude.com/mcp, and two scopes, mcp:read and mcp:write.
One caveat on this whole route, because it is the one that bites six months in. You have moved the problem from procurement to data modelling. The assistant's answers will be exactly as good as the model sitting under them, and MCP scopes are not row-level security: the scope gates the API surface, your warehouse gates which rows a given identity can see. If those two controls are not aligned, a read-only scope will still read everything.
What adoption costs an IT team#
Connecting a hosted MCP server takes hours. The cost is the review, and it is worth naming the specific reviews this use case triggers so they can be scheduled rather than discovered.
Egress review, for every server on this page except Supabase and the two Microsoft packages. All five market and consumer intelligence servers, and both enforcement servers, publish no installable package. They are hosted-only, so every query and whatever context accompanies it leaves your network. Six of 33 servers across the whole catalogue offer a self-hosted path; none of those six is a market-data source.
The four control criteria you will be asked about in the security review are undocumented everywhere. Across all 33 servers, none documents audit logging, access control, data residency or retention. Not one, on any of the four. Undocumented is not the same as absent — these may well be handled and simply not written down — but it means the answer cannot be read off a page, and has to be obtained from the publisher in writing and then written into the contract. Budget for that correspondence.
Open endpoints need a decision, not a reflex. Five of the 33 servers accepted an unauthenticated connection when we probed on 19 September 2026, and four of those five are relevant here: MyPlate.food, offerhopper.ai, Argus HQ and RecallRadar. No credential is genuinely less work — nothing to provision, nothing to rotate, no secret in a client config. It also means no per-user identity, so tool calls cannot be attributed to a person. For a public reference feed that is usually acceptable. Decide it explicitly rather than inheriting it.
For the authenticated ones you are buying partly blind. We could not enumerate tools on fmcg.network or January AI without credentials, and neither declares a repository in its registry record — nor do MyPlate.food, offerhopper.ai or RecallRadar. There is no way to read the tool descriptions before you hold an account. Plan a short spike: obtain a trial credential, run a tools/list, read the descriptions and argument schemas, and only then estimate the integration.
Two publisher-diligence notes worth an email each. Argus HQ's namespace is verified to argushq.ai and its endpoint sits on argushq.ai, which is consistent — but the repository declared in its registry record is github.com/andrewjgaber-commits/empire-distribution, an unrelated personal account, and we could not determine a licence. That is a question, not a disqualification. Separately, every namespace verification on this page is a check we ran against registry records on 19 September 2026; treat it as a fact about that date rather than a fact about today.