I’ve watched buyers send the same RFQ to twelve factories and hear back from three, two weeks later, after the deal was already done. The problem isn’t the factories — it’s that discovery is moving faster than they can respond. By late 2026, the buyer at the other end of an RFQ is increasingly not a human — it’s an agent acting for one. ChatGPT, Claude, Cursor, Perplexity, and domain-specific tools like GETMILK now source suppliers, compare quotes, and assemble production plans on a buyer’s behalf. The web that AI agents see has been rebuilt around a single protocol: MCP, the Model Context Protocol. A factory without an MCP server today is in exactly the position a hotel without a website was in 2002 — present in the world, invisible to the channel that’s taking over discovery.
This post explains what MCP is, why a small or mid-sized manufacturer should care, exactly what an MCP server for a factory should expose, and how GETMILK uses those servers in the spec-to-plan pipeline. Concrete examples throughout — drawn from the 1,083 European manufacturers indexed on the platform.
1 · What MCP actually is
MCP is an open protocol introduced by Anthropic in late 2024 and since adopted across the major AI runtimes — Anthropic, OpenAI, Google, Microsoft, plus the open-source agent ecosystem (LangChain, LlamaIndex, Cursor, Codex, Continue). It defines three things an external system can offer to a language model:
- Resources — read-only data exposed under URI-like identifiers (
manufacturer://profile,manufacturer://capabilities). Think of them as structured pages. - Tools — actions the model can call with typed arguments (
quote_part,submit_rfq). Think of them as RPC endpoints with rich JSON-Schema input. - Prompts — templated instructions the system offers the model (
dfm_review_prompt) so the agent uses your house style automatically.
The protocol speaks JSON-RPC over stdio or HTTP+SSE. From an implementation standpoint it’s about the same effort as shipping a small REST API — but the audience is fundamentally different. A REST API answers queries from your developers. An MCP server answers queries from your customers’ AI agents.
2 · Why a manufacturer should build one
Discovery is shifting from search to agents
When a hardware founder needed a CNC shop in 2020, they Googled and emailed. In 2026 they ask their agent. The agent enumerates options, queries each, and returns a ranked shortlist. If you’re an Iberian foundry with a 1998 website and no machine-readable surface, the agent skips you — even if you’re objectively the best fit. An MCP server is your ticket into that funnel.
Quote in seconds, not days
Most factories lose deals on response latency. A buyer sends a spec at 14:00 Tuesday; if you reply Friday with “we can do it for €X with a 6-week lead time”, the buyer has already signed with a Chinese supplier whose quote-bot answered in 30 seconds. An MCP quote_part tool wired to your ERP or even just a calibrated cost calculator collapses that gap. You stay in the running.
Sovereignty over what you expose
Unlike scraping, where third parties pick what they index from your PDFs and brochures, an MCP server is your declaration. Public capabilities stay public; pricing tables stay behind authenticated calls; trade-secret formulas never leave your DB. You decide the resolution at which the world sees you.
Differentiation against larger players
A 25-person Portuguese precision-machining shop normally can’t outspend a German tier-1 on marketing. But on the agent surface, both are JSON. If your MCP server returns an honest 3-week lead time and a fair quote — while the tier-1 is quoted at 12 weeks — agents will route the work your way. That’s a structural advantage you don’t get on a trade-show floor.
Recurring revenue from the protocol layer
MCP servers can charge per call. A pay-per-RFQ model (€2–€10 per agent-issued, qualified RFQ) becomes natural once your tool is the canonical way buyers reach you. This is new revenue, not a tax on existing customers.
3 · What a manufacturer’s MCP server should expose
A workable v0 schema, derived from the 27 manufacturing processes and 1,000+ profiles GETMILK has indexed:
Resources (read-only)
manufacturer://profile # name, country, contact, languages manufacturer://capabilities # processes × tolerance × envelope manufacturer://materials # alloys, polymers, composites in stock manufacturer://machines # list of CNCs/presses/lasers manufacturer://leadtimes # current backlog by process manufacturer://certifications # ISO 9001, AS9100, IATF 16949, … manufacturer://references # case studies, public clients manufacturer://policies # MOQ floor, NDA template, payment terms
Tools (actions the agent calls)
check_capability({ process, material, tolerance, max_dim_mm })
// → { fit: yes|partial|no, reason }
quote_part({ spec_url, qty, target_date })
// → { price_per_unit, lead_time_weeks,
// assumptions, valid_until }
submit_rfq({ spec_url, qty, target_date, contact })
// → { rfq_id, internal_owner, sla_hrs }
book_capacity_window({ process, start, end, est_qty })
// → { slot_id, soft_hold_until }
request_sample({ spec_url, ship_to })
// → { sample_id, est_ship_date }
get_dfm_feedback({ spec_url }) // → { issues[], suggested_changes[] }Prompts (your house style, lifted into the agent)
quote_request_template # preferred RFQ format dfm_review_prompt # design-for-manufacturing checklist material_substitution_prompt # when a customer asks for X but Y is faster
4 · How GETMILK plugs into your MCP server
GETMILK’s pipeline today walks a spec through three stages: analyse (extract a structured product brief from whatever the customer dropped), breakdown (map to processes and materials, propose sourcing options), and plan(build a 90-day action plan with budget and contacts). The breakdown stage is where MCP servers transform the output from “here are some EU manufacturers who’ll probably fit” to “Riopele has 800 kg of Nm 30/2 cotton in stock, can spin from week 18, €5.40/kg at this volume.”
The flow, end-to-end
- Customer drops a spec on getmilk.tech — sketch, photo, link, or a paragraph of notes.
- GETMILK’s analyse step extracts a structured brief (processes, materials, tolerances, target volume).
- The breakdown step queries the internal manufacturer graph (1,083 EU shops as of April 2026) for candidates that match on process and country preference.
- For each candidate that publishes an MCP endpoint, GETMILK calls
check_capabilityin parallel — typically 8–15 calls completing in under 4 seconds — to filter real fits from nominal matches. - For the top 3 fits per process, GETMILK calls
quote_partwith the spec URL — getting back a real cost band, real lead time, and validity date stamped by the manufacturer’s own system. - The action-plan step weaves these into a concrete plan: who to contact, what to send, when to expect first-article, what it costs.
- When the customer presses “Submit RFQ”, GETMILK calls
submit_rfqon each chosen MCP server. The manufacturer’s CRM/email/Slack receives a pre-formatted, deduplicated request — no copy-paste, no language barrier, no missing fields.
5 · Real walkthrough — TEXFIO conductive yarn
Take the textile spec recently launched on the platform: TEXFIO — Washable PEDOT:PSS Conductive Yarn for Smart Apparel. The customer wants 50 km of conductive yarn coated for ≥50 wash cycles, base yarn Nm 30/2 cotton.
Without MCP, GETMILK returns a thoughtful but hand-wavy plan: “Riopele or Polopiqué likely fit; expect ~€5–7/kg; ~12 weeks lead time.” Useful, not actionable.
With MCP servers running at each Vale do Ave mill, the breakdown step instead resolves to:
riopele://check_capability({
process: "ring_spinning",
material: "combed_cotton_Nm30_2",
finishing: "PEDOT:PSS_dip_coat"
}) → { fit: "partial", reason: "spinning yes, coating via CeNTI partner" }
riopele://quote_part({
spec_url: "getmilk.tech/spec/a019c21f.../export",
qty_kg: 50,
target_date: "2026-09-15"
}) → {
price_per_kg: "€5.40",
lead_time_weeks: 8,
assumptions: ["base yarn from stock", "coating subcontracted to CeNTI"],
valid_until: "2026-05-13"
}
inovafil://quote_part({…}) → { price_per_kg: "€4.95", lead_time_weeks: 11, … }
polopique://quote_part({…}) → { price_per_kg: "€5.85", lead_time_weeks: 6, … }GETMILK now writes a sourcing plan that names Riopele, Inovafil, and Polopiqué with their own numbers — not generic ranges. The customer sees three real choices, each with a per-kg price stamped “valid until 13 May 2026”. They click “Submit RFQ” on the fastest-lead-time option; Polopiqué’s sales lead gets a fully-formed RFQ in their CRM 90 seconds after the customer first dropped a paragraph of notes on getmilk.tech.
6 · Where this lands first across the European graph
The platform’s 1,083-manufacturer registry is concentrated in a handful of clusters where MCP adoption will compound quickest:
- Vale do Ave (Portugal) — textile spinning, knitting, dyeing, finishing. ~40 mid-sized mills with existing ERP systems (PHC Software, Primavera) that could wrap an MCP layer in a week.
- Lombardia + Veneto (Italy) — precision CNC, sheet metal, leather goods. 118 indexed; the German tier-1s already pull from here, MCP would let smaller buyers in.
- Baden-Württemberg (Germany) — 81 indexed; mostly automotive-grade CNC and EMS. Slowest to adopt due to existing EDI investments, but largest deal sizes once they do.
- Stara Zagora + Plovdiv (Bulgaria), Łódź (Poland)— emerging composites, drone parts, and EMS. Hungry for agent-driven discovery because they don’t have the sales team to chase EU-wide.
7 · How a manufacturer actually ships one this quarter
- Inventory what you already have. Capability list in a Word doc, machine list in Excel, lead times in sales-rep memory. Most of v0 is normalising existing data, not generating new data.
- Pick a runtime. Anthropic and the community ship MCP SDKs in TypeScript and Python. A Node.js implementation behind a small Vercel function is fewer than 300 lines.
- Wire to your source of truth.If you have an ERP — even a 2008 PHC install on Postgres — write a thin read-replica view and have the MCP tools query that. If you don’t have an ERP, a Google Sheet works for v0.
- Publish at
your-domain.com/.well-known/mcp.jsonso agents can discover it without you registering anywhere. GETMILK auto-detects MCP endpoints when crawling supplier websites. - Instrument usage.Log every tool call with spec URL, agent identifier, result. Within a week you’ll see which tools matter.
Closing
The web reorganises every fifteen years. Search reorganised it around pages; mobile reorganised it around apps; agents are reorganising it around tools. For a manufacturer, this isn’t the abstract change it sounds like — it’s the difference between being a name an agent enumerates and being a name no agent has ever heard of. The work to close that gap is small; the cost of staying on the wrong side of it grows weekly. Build the server. Stay findable.