Sphere Quarterly · Issue 01 · Agents
Your site has an API for humans. Now it needs one for agents.
Agent-Ready Sites — Sphere Quarterly · Issue 01
Your site has an API for humans (HTML). In 2026 it needs one for agents (MCP). A field guide to building marketing sites that LLM agents can actually call, by the team that ships the reference implementation.
- Anton MaciusField CTO
In this article
- The crawl is dying. The call is arriving.
- Request, route, respond. The wire flow is small.
- Nine tools, six resources, one front door.
- Same content. Four readers. Four front doors.
- What did the agents say about us last quarter?
- Why this, why now: every platform shipped agents.
- This piece is honest about the line.
Every major platform shipped agentic browsing in 2024–2025. Claude calls tools. ChatGPT calls Apps. Copilot calls connectors. The agents are arriving, and they don’t read HTML — they read structure. Sites that publish an MCP endpoint get called. The rest are left to be interpreted.
If you only read one box on this page.
For thirty years, websites have served exactly one client: a person sitting behind a browser. In 2026 a second client showed up — AI agents. Claude, ChatGPT, Copilot, Cursor, custom GPTs, the next thousand. They don’t need pictures or padding. They need structure.
Today they fake it by scraping the HTML and guessing what your site does. They guess wrong frequently. They miss the “Book a demo” button because it’s an aria-label inside a <button> inside a modal. They misstate your pricing. They send users to your competitors’ pages by accident.
Sphere gives every site a second front door. Every site publishes an MCP endpoint — Model Context Protocol (opens in new tab), the open standard Anthropic open-sourced in late 2024 and every major AI platform has adopted. Agents stop scraping. They start calling. The site exposes typed tools and signed resources. Every interaction is auditable.
The result: agents that act on the user’s behalf find the right page, book the right demo, pull the right blog post, and never hallucinate your pricing again. The site stays a website for humans. Underneath, it’s an API for everything else.
A user asks Claude: “Book me a demo of that AI compliance product.”
Claude scrapes the site, guesses which form to fill, and gets the timezone wrong half the time.
Claude calls book_demo on the site’s MCP endpoint. The booking lands clean with timezone, owner, source.
A custom GPT helps an analyst summarize three competing vendors.
It pulls cached crawls and hallucinates feature support.
It calls list_products, get_blog_post, search_content. Returns verbatim, with citations.
A regulator asks “what does your AI tell agents about us?”
Nobody knows what crawlers cached or how a model paraphrased it.
Every MCP call is signed and logged. The full transcript is queryable.
The crawl is dying. The call is arriving.
The crawl is a thirty-year-old workaround. Search engines invented it because there was no other way for a piece of software to learn what a site contained. The arrangement was always uneasy — sites paid bandwidth to serve cached snapshots, and crawlers paid in cycles to parse HTML that was never meant to be read by machines. Both sides muddled through because there was no alternative.
In 2026 there is. Anthropic open-sourced Model Context Protocol (opens in new tab) in November 2024. Within a year, OpenAI, Microsoft, Google, Cursor, Replit, and every serious agentic product had implemented it. MCP is the wire format an LLM uses to call a tool — structured, typed, authenticated, auditable. It is the HTTP of the agentic web.
Sites that publish an MCP endpoint stop being parsed and start being called. The same engineering effort that wrote the public-facing booking page now answers book_demo(). The same content that lives in the CMS answers get_blog_post(slug). Zero duplicated work, dramatically higher fidelity, signed by default.
Request, route, respond. The wire flow is small.
The protocol is deliberately small. An MCP server advertises a list of tools (callable functions with typed inputs and outputs) and resources (addressable URIs the agent can fetch). A client sends a JSON-RPC request; the server validates, executes, and returns a typed result. The framing is identical to what any modern API uses, with two material differences: tools are discoverable (the agent can ask the server what it can do), and resources are content-addressed (an agent can cite, not just summarize).
For a marketing site, the right surface is small: a search tool over your content, a list of blog posts, a fetch-by-slug for full bodies, a booking tool, a contact-form submitter, a newsletter subscribe. Sphere’s reference implementation exposes nine tools and four resource families. Total surface area: ~600 lines of TypeScript. One afternoon of work for a marketing site that’s now agent-readable for the next decade.
Nine tools, six resources, one front door.
A surface this small is the point. The temptation when designing an agent-readable site is to expose everything. Resist it. Agents work best with a sharply curated tool list — fewer choices, clearer names, typed inputs that prevent stupid mistakes. The marketing site is not the CRM; the CRM has its own MCP server with its own scope.
The right tools cover three jobs: read what we’ve published (search, list posts, fetch a post, list products), act on the visitor’s behalf (book a demo, submit a contact, subscribe to the newsletter, register a referral), and recall what the visitor has chosen to share (their engram, scoped by consent). That’s nine. Most sites will need fewer.
Same content. Four readers. Four front doors.
Once a site is agent-readable, the rendering surface multiplies. The same blog post is, in different contexts, an HTML page for humans, a JSON-RPC response for agents, a markdown body for traditional indexers, and a set of engrams an AI assistant can ground a longer conversation against. The content does not have to be authored four times. It has to be authored once, in a structure that can be projected four ways.
This is the architecture every CMS will be rebuilt around between 2026 and 2028. The marketing teams that get there first will show up in agent answers the way the early SEO movers showed up in Google answers — quietly dominant, with budgets locked in.
What did the agents say about us last quarter?
EU AI Act Article 13 (opens in new tab) requires AI deployers to provide transparency about the information their systems use. When an agent answers a buyer’s question about your product, the agent is a deployer of AI and your site is part of its information chain. If a regulator asks “what did your AI tell users about this company last quarter” — from a competitor’s point of view, that’s a discovery request waiting to happen — the answer is somewhere in millions of crawls and chat logs.
An MCP endpoint pre-stages the answer. Every call is logged, signed, parent-linked. What the agent asked, what your site returned, the canonical URLs cited. Disputes about misrepresentation become a query, not an investigation.
Why this, why now: every platform shipped agents.
Anthropic open-sourced MCP (opens in new tab) in November 2024. Within six months it was the de-facto interface for Claude Desktop, Cursor, Zed, dozens of agent frameworks. OpenAI shipped Apps and Operator. Microsoft shipped Copilot connectors. Google shipped Gemini agents. Salesforce shipped Agentforce (opens in new tab). The industry standardised faster than any prior protocol — faster than REST, faster than GraphQL, faster than OAuth.
The buyer behavior is shifting now. The same pattern that played out with Google in 2001–2005 — sites that adopted SEO early dominated for a decade — is playing out for agentic discovery in 2026–2028. The cost of adding an MCP endpoint is one afternoon. The cost of not having one in two years is being absent from every agentic recommendation engine on the open web.
The regulatory tailwind is real. EU AI Act (opens in new tab) general-purpose-AI obligations are applicable from 2 August 2025; high-risk system obligations from 2 August 2026. Article 13 transparency and Article 26 deployer-logging requirements make signed, queryable agent interactions the path of least friction. Sites that publish an MCP endpoint with the ledger turned on satisfy the documentation requirement by construction. Sites that don’t are reconstructing logs from CDN headers.
Cited
- Model Context Protocol — modelcontextprotocol.io (opens in new tab) (Anthropic, Nov 2024)
- EU AI Act Reg. (EU) 2024/1689 — EUR-Lex (opens in new tab); Arts. 13, 26, 50
- GDPR Reg. (EU) 2016/679 — EUR-Lex (opens in new tab); Art. 30 records of processing
- Agentic ecosystem references — Salesforce Agentforce (opens in new tab), OpenAI Apps, Microsoft Copilot connectors, Google Gemini agents.
This piece is honest about the line.
Sphere’s marketing site is the proof. sphere-iq.com/agent (opens in new tab) documents the live MCP endpoint at /api/mcp. We eat our own dog food. The list below separates what runs today from what ships in the next two quarters.
Nine controls running now. Five extensions in the next two quarters.
Where we are, where we’re going
- Live MCP endpoint —
sphere-iq.com/api/mcp, MCP spec 2025-06-18. - Nine documented tools — search, list, fetch, book, submit, subscribe, refer, recall.
- Four resource families — blog, platform, trust, manifesto.
- Origin allowlist + per-IP rate limit at the edge.
- Audit ledger — every call signed and parent-linked.
- Consent-gated
recall_engrams— visitor key required. llms.txt+llms-full.txt— agent-discoverable site map.- Public agent docs at
/agentwith copy-paste setup for Claude Desktop, Cursor. - ~600 lines of TypeScript — small enough to fork into your own site this week.
- OAuth-scoped tools — agents authenticate as a known partner before calling commercial endpoints.
- Streaming tool results — long-running operations return progress events.
- Per-agent rate budgets — preferred-partner agents get higher ceilings, anonymous agents stay throttled.
- Auto-published
llms-full.txt— regenerated when the CMS publishes. - Agent transcript export — one-click GDPR Art. 30 record for every agent interaction.
More to read
Generative AI integration explained: the patterns, architecture, security controls and rollout steps for adding GenAI to the systems your business already runs.
Custom AI development explained: when it beats off-the-shelf AI, what a build involves, and how in-house and outsourced teams compare on speed and risk.
A shortlist of top AI development companies for 2026, grouped by firm type, with a transparent selection method and advice on choosing the right partner.
Generative AI use cases in financial services by function: what each does, the data it needs, the controls regulators expect, and how to measure value.