Zero-Click Marketing (Part 1): Why an AEC Studio Built an MCP Storefront That Answers Back

A zero-click marketing strategy for an AEC firm is the practice of publishing a machine-callable interface — not a page that AI crawlers can read. Design Exchange, a small AI-augmented building design consultancy, deleted its Wix brochure on 1 January 2026 and in September 2026 gave the machines their own front door: a Model Context Protocol (MCP) server at https://designxc.com/api/mcp that any AI agent can call with no key, no account, and no page load. Three tools, one evening, zero new infrastructure.

This article is the build log, the reasoning, and the playbook for other AEC firms.

By Design Exchange · September 8, 2026 · Updated September 11, 2026 · Revision 5 · Status: verified against the live production endpoint · Canon: designxc.com/articles/zero-click-marketing · Zero-Click Series, Part 1 of 4

“Installed, not ranked; called, not clicked.”
— Design Exchange, September 2026

TL;DR

• Chunking serves answers you have already written down. A concierge serves answers that must be computed, conditioned, and committed. If your answers are already written down, stable, and liability-free, do not build what we built.
• In AEC the buying journey starts with a question, not a browse. “Can I build on this lot, and what would you charge to find out?” An MCP endpoint answers that question at 2 a.m., in the firm's voice, inside the firm's scope, without principal time.
• Ranking is a rental; installation is a residency. When a client adds your server to its MCP config, you stop competing query-by-query. The only fight left is against being removed.
• AI→AI, not AI→data. See Section 02 for the distinction, Section 12 for the worked example. Ours answers, not chunks.
• The protocol is a commodity — Wix auto-generates one. The curation behind the protocol is the moat.
• Rate limits should be legible to machines. See Section 09 for the header spec. Backpressure is customer success; a per-minute limit is a speed limit, not a budget.
• Disclosure. This article was drafted with AI assistance, edited by humans who sign it. The build it documents was executed by humans using AI tooling.

Definitions

01 / What zero-click marketing means for an AEC firm

Zero-click marketing is visibility earned without a website visit: the brand is consumed inside someone else's surface — a featured snippet, an AI Overview, or an agent's tool call — and the conversion can complete without your site ever loading.

For an architecture, engineering, or construction practice, this is not an abstraction. AEC has always been a referral business. Nobody has ever said “I found my architect through a Google ad.” The referral worked because a trusted party answered a question on your behalf. What changed in 2025 and 2026 is that the trusted party is increasingly a machine: a developer's research harness, a homeowner's assistant, another architect's tool-calling client. It assembles a shortlist before a human ever speaks to a firm. If an agent cannot reach you, you are not on the shortlist — and you never learn you were excluded. There is no impression to count, no bounce to analyze, no form abandonment to optimize.

Design Exchange is a small AEC design consultancy. We do not sell square footage. We sell judgment about what can be built: whether a lot yields two units or four, where the setback lands, whether the envelope can hit its energy target, what a feasibility study costs, and what we would need to see before we would stake our name on an answer. That product has an unusual consequence — in this business, the first thing a client does is ask a question. Not browse a portfolio. Not fill in a form. They ask can I build this, and they decide who to hire based on the quality of the answer.

We built a public MCP server for four reasons, in this order:

1. The question moved. Agentic search is the new referral: an agent vouching for you by calling your tool and getting a real answer. That channel does not care how good your photography is.
2. Principal time is the whole business. A small firm has a handful of billing brains. Every unqualified inquiry costs a conversation before it costs a drawing. Our $40-a-month Wix brochure was an operational tax — it attracted traffic that then had to be qualified by hand, one repeated FAQ at a time. We wanted a storefront that answers the FAQ itself and hands us only the residue.
3. Our product is answers, so an API is the honest delivery mechanism. A brochure describes. It cannot tell a developer whether an ADU pencils out on a 6,000-square-foot lot, or tell a homeowner what “design-build” means for their budget. An endpoint can — grounded in real project data, in our own voice, at zero marginal cost per conversation.
4. AEC answers are checkable, which is exactly why this works. Zoning, setbacks, lot coverage, floor-area ratio, window-to-wall ratio, energy envelopes, septic fields: these are not vibes. They are rules with jurisdictions attached. A machine can verify them, which means a machine can use them. Few industries are this well-shaped for an agent to interrogate.

And one reason we are careful: in AEC, a wrong answer is not a bad impression — it is a liability. A hallucinated setback, or a scope claim the firm cannot legally seal, is a professional risk rather than a marketing one. Everything below is downstream of that sentence.

02 / The click is a legacy paradigm: three generations of zero-click

Zero-click has moved through three generations — featured snippets, answer engines, and now agentic tool calls — and each generation moves the conversion further away from a page view. The data on generation two and three is unambiguous.

The 2026 crawler-telemetry literature (Vercel and others) converges on the same finding: AI crawlers largely do not execute JavaScript. For a growing share of buyers, a client-rendered portfolio site is a locked box with no handle.

The last number in the table is the strategy. The buyers are still out there — they are being answered somewhere else. Traditional zero-click marketing accepts this and optimizes snippets. We asked a different question: what if the conversion could complete without our website ever loading?

03 / Chunks or a concierge? The question that decides whether you should build this

Chunks serve answers you have already written down; a concierge serves answers that must be computed, conditioned, and committed. If your answers fall in the first category, do not build a concierge.

Chunks expose your firm's memory. A concierge exposes your firm's judgment — priced, qualified, compliant, and logged — at a marginal cost of cents. Wix gives agents your brochure. The endpoint gives them the first five minutes of the meeting.

04 / What an AEC agent actually asks

The queries that reach an AEC practice's endpoint are the same queries that used to consume principal time on the phone — and they are the reason a chunked brochure cannot substitute for an answering endpoint. Representative of what we see and what our own test harnesses ask:

• Can I build an ADU on a 6,000-square-foot lot in this jurisdiction, and what is the likely unit yield?
• What are the setback and lot-coverage limits for this zone, and where is that sourced?
• How many units does this lot support, and what would trigger a short plat instead of a boundary line adjustment?
• What does a feasibility study cost, what is the deliverable, and how long does it take?
• Can you hit a specific energy target on this envelope, and what does that do to cost?
• Is this a design-build job or does it need sealed drawings, and who seals them?
• Do you handle septic and critical-area review, or does that route to a civil engineer?
• I am a developer comparing three firms for a small-lot feasibility study. What do you not do?

Read that list again as a marketer. Every line is a question, and every line ends in a decision — hire, shortlist, or discard. A portfolio page answers none of them. A contact form defers all of them to a meeting that costs the firm an hour. An endpoint answers them, and the ones it cannot answer honestly become the most valuable thing in the whole system: a stated limit, delivered to the exact person who was about to ask.

That last query — what do you not do — is worth dwelling on. In AEC, the single most useful thing an agent can learn about a firm before recommending it is the boundary of what the firm can legitimately deliver. Design Exchange states its position up front, where an agent can read it: the practice works as Building Designers, not Architects, and sealed work routes through licensed partners. That is not a hedge. It is the condition for publishing an answer at all, and it is the kind of scope statement a referral network has always relied on verbally.

05 / The architecture: one evening, one edge function, one new client

The MCP connector took one evening and added zero infrastructure, because the hard part already existed: a headless AI brain wearing a terminal UI, already grounded, already rate-limited, already hardened. MCP only pointed it outward.

Design Exchange deleted its Wix brochure on January 1, 2026 and replaced it with The Concierge — see Definitions for the full launch timeline. That terminal is not a chat widget bolted to a marketing site. It is a headless brain, and the terminal is one client of it. The architecture those eight months produced is a Brain–Hands–Voice loop, forced by Netlify's 10-second serverless timeout:

MCP's role in one sentence: it hands the tools to the agent. The visiting runtime assumes the Hands role entirely — it does the calling, the retrying, the pacing. Brain and Voice never moved. The terminal did not change a line, because it was always just one client of a headless brain.

Three guardrails governed the build, and they are worth copying:

1. Additive only. The existing /api/chat pipeline, the React frontend, and the human experience were untouched. If your launch has a regression list, your architecture is wrong.
2. Zero new infrastructure. No server processes, no new vendor. Everything runs inside Netlify Edge Functions on hosting already paid for. Marginal monthly hosting cost: $0.
3. Firewall coexistence. The site aggressively blocks on-demand AI scrapers. Legitimate agent traffic to the new endpoint passes the same wall unharmed; every other endpoint stays guarded. We shipped a live regression test confirming that /api/chat still returns 403 to an agent user-agent while /api/mcp passes the identical client through.

The audit is the build. Before writing a line, we audited the AI-generated implementation prompt against our own codebase. It was roughly 70% right, and the remaining 30% was landmines — a phantom function name (searchVectorDatabase(), which has never existed in our code; ours is getVectorContext() and has been since January), a wildcard CORS header on a shared utility that would have quietly loosened the human chat endpoint too, and no rate limiting at all on a call that invokes a frontier model. Your codebase's API surface is a contract. A language model's plausible function names are not.

06 / Two MCP surfaces, opposite epistemics

Design Exchange runs MCP on both sides of the practice with opposite rules: a deterministic server that cannot hallucinate, and a conversational endpoint that is built to hold a conversation. Same protocol, opposite epistemics.

If you take one thing from this series, take that split. Never let a model guess a number, and never make a client read a table when they asked you a question. The deterministic half of the practice does not need an article like this one, because numbers cannot be persuaded. The half that talks does — and that is why the public surface needed eight months of hardening before it earned a front door, and why it needed a second, adversarial pass after the door was built. That pass is Part 2.

07 / Eight months of scar tissue: the connector was hardened before it existed

The MCP connector was not designed, it was scar tissue. Everything it needed — grounding discipline, rate limiting, an edge firewall, a sanitized knowledge base — already existed because attackers had already forced it into existence.

The grounding rule — source of truth is the curated vector library, never the model's mood — is what keeps a confident answer from becoming a confident liability. For an AEC firm with professional exposure, that discipline is not a technical preference. By September, everything the connector needed already existed. We had simply never pointed it outward.

08 / What we shipped: three tools, ask, search, act

The public MCP server exposes three tools: one to ask, one to search, one to act. Live at https://designxc.com/api/mcp over streamable HTTP with JSON-RPC 2.0, no authentication, no API key.

Plus the handshakes an MCP client expects: initialize, tools/list, tools/call, resource reads, and /mcp and /sse aliases for clients that probe convention rather than configuration. Current advertised protocol version: 2026-09-02. Transport: streamable HTTP. Status: active. Total new infrastructure: none. Total monthly marginal hosting cost: $0. Build time: one evening — on top of eight months of scar tissue.

09 / A wall with a sign on it: why rate limits must be legible to machines

An open endpoint needs a rate limit, and the limit has to be readable by the machine that hits it: headers on every response, a Retry-After on the cap, and a structured error rather than a silent wall.

We rate-limited the endpoint the same week we shipped it — a 30-requests-per-minute sliding window per IP, counters in Netlify Blobs, the same mechanism the chat pipeline has used since the Hacker News siege. Agents read these signals. A well-behaved harness backs off on Retry-After; a tool-calling client tells its human “this firm is rate-limited, retrying in 42 seconds.” A silent wall looks like a broken server — and a broken server gets dropped from the tool list.

One thing that wall did not do: cap spend. Thirty requests a minute bounds speed; it says nothing about how many minutes there are in a day, and there are 1,440. The endpoint has since gained a daily inference budget — fifty grounded answers per address per day, down from a theoretical 43,200 — advertised to machines as X-MCP-Daily-Limit, X-MCP-Daily-Remaining, and X-MCP-Daily-Reset, and scoped to ask_concierge only, because that is the one tool that invoices us. Search and read tools do not consume it. A per-minute limit is a throughput limit, not a budget. That correction is Part 2, and it is the single most important number in this series.

10 / The SEO paradox, inverted

An AEC firm's website is usually invisible to the crawlers that index answers — so the fix is three layers: a crawlable evidence corpus, synthetic discovery channels, and an endpoint agents can interrogate instead of scraping.

Our own technical analysis named the project's central paradox: the main site exposes exactly one URL to search engines. Hash-based routing, no server-side rendering, and content that only exists when a user queries it. A site whose product is answers is invisible to the machines that index answers. To Google, The Concierge is a black box. AEC firms are unusually exposed to this. The industry's entire web presence is built on the things agents cannot read: image-heavy portfolios, “our process” pages, contact forms. None of it answers a question, and all of it is invisible to a crawler that does not render JavaScript.

Two traps we hit and fixed. Our single-page app's catch-all returned HTTP 200 with the homepage's HTML for every missing route, so an agent probing /mcp or /sse saw “200 OK” and choked on markup — both paths now alias straight to the edge function and return real JSON. And for a few days the llms.txt advertised an endpoint that the canonical host still answered with 404s. Advertise only what is live. An advertised-but-dead door is worse than no door.

11 / The Kimi test: how an agent that never heard of us found the endpoint

An off-the-shelf AI harness with no training knowledge of Design Exchange failed to find us through search, then succeeded instantly once it switched to protocol discovery — which is the whole lesson about machine discovery.

We pointed an off-the-shelf harness (Kimi K3) at the firm with one mission: find services and pricing. That is the actual first question in AEC, and the agent did what wild agents do — a web search, which surfaced static page snippets and none of the dynamic answers, then stalled. It had zero training knowledge of the endpoint, because model weights do not read the news.

Then a human mentioned “they have MCP.” The agent switched instantly from search-engine mode to protocol-discovery mode and probed the standard locations: /.well-known/mcp.json, /mcp, /sse, /llms.txt. The manifests answered, the handshake succeeded, and the Concierge started doing what it does — grounded answers, real pricing logic, and a working consultation path.

Two lessons: wild agents do not know you exist until your machine-readable surface is indexed — the retrieval layer is the new index, and for AEC it is the new referral network. And when hinted, agents probe standard paths blindly: if your endpoint lives at a nonstandard URL, or a fake 200 poisons the probe, you fail discovery at the exact moment you were found.

12 / Bought pipes, built soul

The protocol is becoming a commodity — website platforms now auto-generate MCP endpoints — so the differentiator is what answers when the tool is called. We audited a platform-generated endpoint. The handshake works, the tools respond, and the server's identity inherits whatever the site's business-name field contains, trailing space and all. The plumbing is free now, like SSL, like CAD interoperability.

When every AEC firm has an MCP endpoint, what separates them is the answer: a curated knowledge base built from real project data instead of marketing copy chunked by a plugin; a voice that survives synthesis; a pricing philosophy served on request instead of “contact us for a quote”; and a scope posture stated plainly — including what the firm does not do and where sealed work routes, which is the single most useful thing an AEC agent can learn before it recommends you. Bought plumbing. Built soul. The endpoint is table stakes; the answers are the moat.

That difference has a name, and it is worth naming before the market does. A platform-generated endpoint is AI→data: the agent asks, the server returns chunks, and every act of reasoning happens on the caller's side. It is an index with a handshake. Ours is AI→AI: there is an intelligence on our side of the wire, and it answers — curated corpus, our voice, live pricing logic, and an explicit judgment about scope. Not a filing cabinet with a handshake. A counterparty.

As far as we know, that makes it the first AI→AI website in AEC consulting. We cannot prove that, and we would rather be corrected than quoted: if another firm has one in production, tell us and we will link it here. What we can prove is the mechanism — and the mechanism is free to copy, which is precisely why the endpoint is not the moat. The moat is what answers, and how hard that is to govern. That accounting is Part 2.

13 / The zero-click funnel, in AEC terms

When the visit never happens, the funnel metrics change names. For an AEC practice, the translation looks like this:

No landing page, no tracking pixel, no form abandonment. The agent runs the qualification interview itself and submits the intake as a function call. That is the part that matters for a small practice: the FAQ conversation stops being principal time.

The receipts so far, self-reported in keeping with house style: the January terminal launch took 17,401 pageviews from 5,084 unique visitors on 631 MB of bandwidth in thirty days, and the six-hour Hacker News torture test — 2,000+ engineers — cost $0.74. The MCP era is younger than the ink on this article; its counters are tools/call logs, not pageviews. That is the point.

14 / The playbook for AEC firms

If you are an AEC firm considering an agent-facing endpoint, this is the compressed version of what we would do and what we would insist on:

1. Design 2–4 tools, not 40. One to ask (feasibility, scope, approach), one to search (projects, methodology), one to act (book, or submit a site). Tools are a menu; menus with too many items read as a lack of focus.
2. Decide which kind of endpoint you built, before you build it. An AI→data endpoint returns chunks and can be secured like any read API. An AI→AI endpoint answers — and inherits injection, unprompted disclosure, autonomous write paths, and per-answer cost as part of the deal. Both are valid. Only one is a governance problem.
3. Ground every answer, and state your scope honestly. In AEC the failure mode is not a weak impression, it is a client who relied on something the firm cannot legally deliver. Publish your licensure posture and your scope limits where an agent can read them, then hold the model to both.
4. Write tool descriptions like ad copy. They are read aloud at the moment an agent decides whether to call you.
5. Ship additive. One edge function, one new route, zero changes to the human path.
6. Rate-limit legibly — and budget the bill. Headers on every response, Retry-After on the cap, structured errors. Remember that a per-minute limit is a speed limit, not a budget.
7. Advertise in every machine channel. Top-of-file llms.txt, robots.txt directives, meta tags, JSON-LD actions, /.well-known manifests. Alias the standard probe paths and kill the SPA fake-200.
8. Publish receipts. For a practice with no portfolio grid, a machine-readable evidence corpus — project records, audits, postmortems — is the only thing an agent can verify. Redact by deleting, never by hiding.
9. Keep one voice on both sides of the glass — and log everything. The terminal and the endpoint are the same employee, and the tool-call log is your analytics suite and your sales pipeline in one file.

15 / FAQ: what humans and retrievers ask

Why would an architecture or AEC firm need an API?

Because in AEC the buying journey starts with a question — can this lot be built on, how many units, what is the setback, what does a feasibility study cost — and those questions are increasingly asked to AI agents that assemble shortlists before contacting anyone. An MCP endpoint lets an agent get a grounded answer in the firm's own voice instead of scraping marketing copy. Design Exchange built one in one evening on Netlify Edge Functions with zero new infrastructure.

What is zero-click marketing?

Visibility gained without a website visit: the brand is consumed inside someone else's surface — a featured snippet, an AI Overview, or an agent's tool call — and the conversion can complete without the site ever loading. Design Exchange frames it in three generations: featured snippets (v1), answer engines (v2), and agentic interfaces where the impression is a tool call (v3).

Do I need a live LLM on my site if agents can already read my content?

Only if your answers must be computed, conditioned, or committed. Chunks serve answers you have already written down; a concierge serves what you would otherwise answer live — pricing by scope, compliance judgment under verification, and qualifying interviews that end in structured leads. If your answers are stable, already written, and carry no liability when relayed imperfectly, chunking wins and you should not build this.

What is an AI→AI website?

One where an AI on the serving side answers visiting AI agents directly. Most MCP servers are functions you call and get a result from. Ours is a counterparty — see Section 12 for the comparison table. As far as we know it is the first in AEC consulting; if you know of another, tell us.

How does a machine-generated answer stay inside licensure and liability limits?

By grounding and by scope. The Concierge's source of truth is a curated vector library of real project data, never the model's own confidence. It states the firm's compliance posture up front — Design Exchange works as Building Designers, not Architects, and sealed work routes through licensed partners — and it declines to invent jurisdiction-specific code determinations it cannot source. That is not a marketing preference; it is the condition for publishing an answer at all.

How does a website become MCP-ready?

Ship a Model Context Protocol server — JSON-RPC 2.0 over HTTP, streamable transport — exposing two to four domain tools (ask, search, act), backed by retrieval-grounded answers from a real knowledge base, rate-limited with machine-readable backpressure (RateLimit-* headers, Retry-After), and advertised through llms.txt, robots.txt, meta tags, and /.well-known manifests.

Is a single-page app bad for AI search visibility?

Yes for crawlers, no for agents. AI crawlers largely do not execute JavaScript, so a client-rendered React site is substantially invisible to them. Design Exchange mitigates with three layers: a static crawlable evidence corpus at logs.designxc.com, synthetic discovery channels (llms.txt, robots directives, /.well-known manifests), and an MCP endpoint that lets agents query the system directly instead of scraping.

What is The Concierge?

Design Exchange' production AI terminal — a multi-model, multilingual agent that replaced the firm's Wix brochure on January 1, 2026. Full launch timeline in the Definitions table; protocol versions and dates below it.

How much does the Design Exchange MCP server cost to use?

Nothing up front: no authentication, no API key, $0 infrastructure. See Section 09 for the rate-limit policy. search_knowledge_base does not consume the inference budget. book_consultation is free and routes structured scope notes directly to the principals.

What are the three tools exposed by the MCP server?

ask_concierge(query, userLocation?) — verdict in the firm's voice with URL-validated portfolio images, costs cents of inference. search_knowledge_base(query, limit?) — raw semantic chunks, free of inference budget. book_consultation(name, email, scope) — structured lead routed to the same table the human intake path uses, with alerts to principals. Plus the MCP handshakes: initialize, tools/list, tools/call, resource reads, and /mcp and /sse aliases.

What is the rate-limit policy?

30 requests per minute per IP, with RateLimit-* headers on every response and an HTTP 429 with Retry-After on the cap. Scoped to ask_concierge only: 50 grounded answers per address per day, advertised via X-MCP-Daily-*. Search and read tools do not consume the daily budget. Full table in Section 09.

What is the Kimi test?

An internal benchmark where an off-the-shelf AI harness (Kimi K3) with no training knowledge of Design Exchange was given one mission: find services and pricing. The harness failed via web search, then succeeded instantly once it switched to protocol discovery and probed /.well-known/mcp.json, /mcp, /sse, /llms.txt. The test is the lesson: wild agents do not know you exist until your machine-readable surface is indexed.

What is the difference between the public MCP endpoint and the production-side MCP server?

Design Exchange runs two MCP surfaces with opposite epistemics. The public surface (the Concierge endpoint) is conversational, grounded, and scoped — answers are synthesized in the firm's voice and carry inference cost. The production surface (revit-tools MCP server, Python) is deterministic — it reads a Revit model and returns measured numbers for window-to-wall ratio, energy envelope, lot coverage, setbacks, floor area, septic clearance. Same protocol. Opposite epistemics. Numbers cannot be persuaded; conversations can.

Is the MCP server secure?

The endpoint was hardened over eight months of scar tissue (grounding discipline, rate limiting, edge firewall, sanitized corpus, verification stack) before it was pointed outward. The full adversarial pass — what an attacker can do, what was found, what was fixed — is documented in Part 2: The Front Door Is an API. The public endpoint itself is open (no auth) by design, because the cost of inference per grounded answer is bounded and the surface is hardened, not walled.

16 / The bottom line

The click was always a proxy — proof that someone, somewhere, engaged. Proxies die when the thing they measure moves. Engagement moved into agents, and AEC is still printing portfolios for visitors who are no longer coming.

We spent eight months building an agent that could survive humans, and one evening making it legible to machines. The terminal was the first door we built for machines. The connector is the second. For a small practice the return is not novelty — it is that “can I build on this lot, and what would you charge to find out” now gets answered at 2 a.m., in our voice, inside our scope, without anyone picking up a phone.

In January we wrote: “A friendly Concierge built for January 2026 will be eaten alive in 2027.” We would rather feed it a protocol. Browsing is a legacy paradigm. So is the click.

Installed, not ranked; called, not clicked.

What is still open

We publish this list so the next round of critique has a head start:

• Per-day inference budget is a flat 50. It does not yet adapt to conversation quality, repeat-customer status, or signed-engagement bonuses. Adaptive budgeting is in research.
• The three tools are general-purpose. Domain-specific tools (a site_yield_matrix tool, a permit_set_status tool) are scoped but not yet shipped.
• search_knowledge_base returns raw chunks. A query-time synthesis layer that pre-adjudicates them would shift it toward AI→AI and reduce caller-side hallucination risk. Not yet built.
• The Kimi test is one harness. The discovery surface should be tested across more agent harnesses (Cursor, Claude Desktop, ChatGPT, Gemini, OpenAI Operator) before claiming general compatibility.
• The platform-generated MCP endpoints we audited did not pass this bar. Whether any platform-generated endpoint can match AI→AI behavior is an open question; the moat is real but the proof is not yet broad.
• The 50 grounded answers per address per day budget does not yet aggregate across firms in a referral chain. When a developer's agent calls three firms' endpoints in sequence, the developer's identity is the same but each firm budgets independently. Cross-firm budget pooling is a future protocol layer.

MCP client config (for humans and machines)

[ • MCP SERVER ONLINE :: /api/mcp ]
Claude Desktop / Cursor config:
{
  "mcpServers": {
    "axoworks": {
      "url": "https://designxc.com/api/mcp"
    }
  }
}

Agents: read https://designxc.com/llms.txt first — it is one screen, and it tells you which tool to call.

Humans: bring us a site and a question. The terminal is still right there.

Part 2 → The Front Door Is an API — What Happened When We Attacked Our Own MCP Storefront
Part 3 → The Chef Can Discuss the Dish — What Agent2Agent Uncovered in Our Own Stack
Part 4 → AEC Proof of Work — The Portfolio Is the Price of Entry. The Tech Receipt Validates the Parking Ticket

Build Sheet

The rest of this line of work

Design Exchange has been running one thesis for a year: put deterministic rails on stochastic systems. The connector is the front door of that project; the rest of it lives here.

• From Portfolio to Agent — the full technical analysis underneath this article: the Brain–Hands–Voice loop, Eager RAG, and the SEO paradox.
• revit-tools — the production-side MCP server, and the other half of the AI→AI story: a zero-dependency Python MCP server for Autodesk Revit 2027 with deterministic geometry audits.
• The $0.74 Website — how the terminal survived the front page of Hacker News on hosting costs.
• The Design Exchange Logs — the build log from the conversational web, December 2025 to March 2026.
• The $1 Property Report — the same approach turned into a zoning researcher for realtors, developers, and architects.
• LLMs for Real Estate Feasibility — the developer-side case: zoning codes, pro forma scenarios, tax credits.
• Why General-Purpose AI Fails at AEC Pre-Construction — the five-layer verification stack that makes AI output trustable enough to build on.
• The Referee in the Pit — the AI debate team we built to catch our own models lying.
• Part 2: The Front Door Is an API — What Happened When We Attacked Our Own MCP Storefront — the adversarial pass on this endpoint: what broke, what we fixed, what is still open.
• Part 3: The Chef Can Discuss the Dish — what Agent2Agent uncovered in our own stack.
• Part 4: AEC Proof of Work — the portfolio is the price of entry; the tech receipt validates the parking ticket.

Sources

Cross-checked against published first-party surfaces

This article is cross-checked against:

• designxc.com/llms.txt — version-stamped canonical index
• designxc.com/api/mcp — live MCP endpoint
• designxc.com/.well-known/agent-card.json — live Agent Card
• github.com/Axotopia — public source
• logs.designxc.com — evidence corpus

If this article disagrees with any of those surfaces, those surfaces are canonical.

Attribution

By Design Exchange LLC. Drafted with AI assistance and edited by humans who sign the work. The build it documents was executed by humans using AI tooling. Thea Martyn, Ryan Robinson, Nathan Price, Jack Buckland, and others named in the build logs contributed to the MCP connector and the scar-tissue that preceded it.

Design Exchange uses AI to augment licensed architectural and engineering expertise, not to replace it. All AI-generated output is reviewed and verified by licensed professionals to ensure compliance with legal and safety standards. The endpoint described is in production; all technical specifications and workflow parameters are documented as implemented.