Enhance customer experiences with Apollo GraphOS


<p class="wp-block-paragraph">Apollo’s annual conference, and the world&#8217;s largest GraphQL event, returns to San Francisco in October. <a href="http://www.apollosummit.dev">Apollo Summit 2026</a> runs October 6 to 8, 2026, at <a href="https://themidwaysf.com/">The Midway</a>, with sessions and workshops on how a well-governed graph turns into the entry point for agents and applications alike.</p>



<figure class="wp-block-image size-large"><a href="https://apollosummit.dev/"><img loading="lazy" decoding="async" width="1024" height="538" src="https://wp.apollographql.com/wp-content/uploads/2026/07/Speaker-Ann-Social-1200x630-1-1024x538.jpg" alt="" class="wp-image-25647" srcset="https://wp.apollographql.com/wp-content/uploads/2026/07/Speaker-Ann-Social-1200x630-1-1024x538.jpg 1024w, https://wp.apollographql.com/wp-content/uploads/2026/07/Speaker-Ann-Social-1200x630-1-300x158.jpg 300w, https://wp.apollographql.com/wp-content/uploads/2026/07/Speaker-Ann-Social-1200x630-1-768x403.jpg 768w, https://wp.apollographql.com/wp-content/uploads/2026/07/Speaker-Ann-Social-1200x630-1.jpg 1200w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></a></figure>



<p class="wp-block-paragraph">Apollo has always worked on the difficult middle layer between backend systems and the clients that need to access them, from the early Meteor days through GraphQL’s origins as a way to fetch data for web and mobile apps, and now as the layer that connects AI agents to the systems they need to access in order to do real work. If you own a GraphQL implementation, or you are weighing how your API strategy fits into your AI roadmap, this is where the practical patterns get worked out in public.</p>



<h2 class="wp-block-heading"><a href="#meet-the-speakers" aria-hidden="true" class="aal_anchor" id="meet-the-speakers"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a><strong>Meet the Speakers</strong></h2>



<p class="wp-block-paragraph">Your APIs were built for apps. What happens when agents want in? These five speakers have already shipped an answer, and they’ll show their work. Stay tuned for more speakers to be announced soon.</p>



<ul class="wp-block-list">
<li><a href="https://www.linkedin.com/in/jonathanwondrusch/">Jonathan Wondrusch</a>, Staff Software Engineer at Block, argues your company already has a graph, hiding in Confluence pages and a hundred half-documented REST endpoints. He shows how to build an API schema that ships product to customers and becomes the context layer your agents reason against. The surprise is that it makes the business legible to itself.</li>



<li><a href="https://www.linkedin.com/in/clarice-abreu/">Clarice Abreu</a>, Senior Software Engineer at Brex, opened GraphOS and found 35% of their federated graph was dead code. Rather than spend months cleaning 60+ subgraphs by hand, she built an agent that deletes, validates, and opens the PR. It has removed over 100,000 lines, and not one customer noticed.</li>



<li><a href="https://www.linkedin.com/in/gijs-martens/">Gijs Martens</a>, Platform Engineer, and <a href="https://www.linkedin.com/in/bartvanlakwijk/">Bart van Lakwijk</a>, Manager Web Platform at Coolblue, routed all production traffic through their graph in week one. The cutover was the easy part. Winning over a &#8220;just do it&#8221; business meant treating the graph as a utility, not a religion, and never blocking a team. No mandate. The graph sold itself.</li>



<li><a href="https://www.linkedin.com/in/sgorla/">Sharon Gorla</a>, Lead Engineer at Starbucks, has a rule for what survives a rewrite: business semantics and shared contracts, never implementation patterns. As coding agents rebuild whole systems, the gap people miss isn&#8217;t MCP or agent frameworks. It&#8217;s alignment. Clients come and go. Shared nouns and verbs are the durable layer.</li>
</ul>



<h2 class="wp-block-heading"><a href="#why-this-matters-now" aria-hidden="true" class="aal_anchor" id="why-this-matters-now"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a><strong>Why this matters now</strong></h2>



<p class="wp-block-paragraph">Organizations are often pulled in opposing directions: deploying agents quickly vs. keeping control over what they can do. Teams that move fast often hand agents broad access and lose visibility. Teams that lock things down slow every project to a crawl. Summit 2026 is built around a different premise: speed and control are not a tradeoff when the graph sits between your agents and your systems.</p>



<p class="wp-block-paragraph">A self-describing graph gives agents declared, scoped access to exactly the tools they need and nothing more. It reduces token costs by handing agents the right context for each action, enforces field-level security, and improves success rates by making your existing APIs discoverable. The sessions show how teams are already doing this in production.</p>



<h2 class="wp-block-heading"><a href="#what-youll-learn" aria-hidden="true" class="aal_anchor" id="what-youll-learn"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a><strong>What you&#8217;ll learn</strong></h2>



<p class="wp-block-paragraph">The agenda is organized around four themes that map to the decisions facing platform and AI teams right now.</p>



<ul class="wp-block-list">
<li><strong>Agent infrastructure.</strong> Design the governed entry point between agents and enterprise systems. Build a context layer that improves task success rates while keeping costs down.</li>



<li><strong>Governance and security.</strong> Manage agent identity, start from zero-rights defaults, enforce field-level access policies, and handle credentials in a single layer. Give both developers and non-developers a trusted platform to build on.</li>



<li><strong>Agentic API access.</strong> Connect REST APIs, SQL data sources, MCP servers, and third-party systems into your context layer, whether or not they already live in your graph.</li>



<li><strong>GraphQL and GraphOS.</strong> For teams running GraphQL in production today: the latest in Apollo Federation, Apollo Router optimization, schema governance, Apollo Connectors, and the Apollo MCP Server.</li>
</ul>



<h2 class="wp-block-heading"><a href="#who-youll-meet" aria-hidden="true" class="aal_anchor" id="who-youll-meet"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a><strong>Who you&#8217;ll meet</strong></h2>



<p class="wp-block-paragraph">Summit draws platform engineers, AI infrastructure leads, developers, and technical leaders who are building API infrastructure for both applications and agents. It is also where the broader GraphQL community gathers, including the people who maintain the specifications and tools the industry runs on. You will leave with patterns you can apply and a set of peers solving the same problems.</p>



<h2 class="wp-block-heading"><a href="#what-to-expect" aria-hidden="true" class="aal_anchor" id="what-to-expect"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a><strong>What to expect</strong></h2>



<p class="wp-block-paragraph">The format balances learning and doing. Expect keynotes, customer talks, hands-on technical workshops, and time to meet the experts and ask your questions. You can hear about how some of the fastest-moving companies are building agents, learn architectural patterns to apply in your own API strategy, and try it out in a workshop, all in just a few days.</p>



<p class="wp-block-paragraph">Here is the shape of the conference:</p>



<ul class="wp-block-list">
<li><strong>Tuesday, October 6.</strong> Arrivals, registration, and a welcome reception from 5 to 7pm.</li>



<li><strong>Wednesday, October 7.</strong> Sessions and an evening event, 9am to 9pm.</li>



<li><strong>Thursday, October 8.</strong> Sessions and departures, 9am to 3pm.</li>
</ul>



<h2 class="wp-block-heading"><a href="#pricing-and-registration" aria-hidden="true" class="aal_anchor" id="pricing-and-registration"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a><strong>Pricing and registration</strong></h2>



<p class="wp-block-paragraph">Registration is now open. Get your Summit pass for $399, including breakfast, lunch, and unlimited coffee and snacks across the event. Dietary preferences selectable at sign-up.</p>



<p class="wp-block-paragraph">October 6 to 8, 2026 at The Midway, San Francisco.</p>



<p class="wp-block-paragraph"><a href="https://apollosummit.dev/">Register now.</a></p>


Apollo Summit 2026: Turn Your API Platform Into Your AI Platform


<p class="wp-block-paragraph">GraphQL is in Gartner’s Trough of Disillusionment in the <a href="https://www.apollographql.com/resources/gartner-hype-cycle-for-apis-2026">2026 Gartner® Hype Cycle™ for APIs</a>. Good.</p>



<p class="wp-block-paragraph">The trough is what happens when expectations set during the hype phase meet the reality of actual production requirements. People stop hyping and start asking the key questions: Does this thing work? Does it solve a problem we actually have? Can we operate it in production without making everyone miserable?</p>



<p class="wp-block-paragraph">Those are the right questions, and any proposed platform tech has to answer them.</p>



<h2 class="wp-block-heading"><a href="#graphql-adoption-vs-graphql-readiness" aria-hidden="true" class="aal_anchor" id="graphql-adoption-vs-graphql-readiness"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>GraphQL adoption vs. GraphQL “readiness”</h2>



<p class="wp-block-paragraph">With GraphQL in the trough, Gartner notes this observation:</p>



<blockquote class="wp-block-quote is-layout-flow wp-block-quote-is-layout-flow">
<p class="wp-block-paragraph">“Despite the potential, GraphQL adoption for AI and agentic workflows have not translated to increased adoption due to its limited penetration in key agentic use cases.”</p>
</blockquote>



<p class="wp-block-paragraph">That strikes us as basically right. MCP and other agent-oriented protocols arrived with an obvious, legible connection to AI. GraphQL, like other mature technologies, is adapting. Most companies adopted it years earlier to solve the problems of the day: frontend development, service integration, data aggregation, and the general difficulty of getting a large organization’s systems to behave as if they belong to the same company.</p>



<p class="wp-block-paragraph">But we think the sequencing matters. Are agents causing companies to adopt GraphQL? There’s some early signal that they are. But a more immediately interesting question is whether companies that already adopted GraphQL are better prepared to put agents into production. The answer is an unequivocal “yes.” There is another interesting signal in the Gartner report. GraphQL is in the trough, but Gartner forecasts that it will reach the Plateau of Productivity in less than two years.  That’s sooner than MCP, gRPC, and AI Agent Communication Protocols. In our read, that tells us something about maturity. GraphQL is no longer new enough to receive credit for its potential. It has to earn credit for being useful. That is a healthier place for infrastructure technology to be.</p>



<p class="wp-block-paragraph">It helps that GraphQL is not a speculative protocol waiting for its first serious workload. It is a <a href="https://spec.graphql.org/">Linux Foundation-backed open standard</a> that has been used at enterprise scale for years. This matters when you are choosing infrastructure that could outlive several executive teams, architecture strategies, and AI hype cycles. Proprietary protocols can disappear when their owner’s priorities change. Open standards have a chance to compound.</p>



<p class="wp-block-paragraph">In our view, Gartner’s own description of GraphQL makes this case more plainly. The report describes its capacity to serve as:</p>



<blockquote class="wp-block-quote is-layout-flow wp-block-quote-is-layout-flow">
<p class="wp-block-paragraph">“A unified enterprise data model, decoupled from the underlying service implementations. This shared schema fosters a common understanding of business data across teams, streamlines integration efforts, and accelerates the delivery of new features and digital products. As GraphQL adoption grows, it also supports advanced use cases, such as enterprisewide orchestration and AI integration, further amplifying its business value.”</p>
</blockquote>



<p class="wp-block-paragraph">You could remove “AI integration” from that paragraph and it would still describe why many organizations adopted GraphQL in the first place. Put it back in, though, and the connection becomes pretty obvious. Agents need to find data, understand how it relates, retrieve what they are allowed to retrieve, and take actions across systems. A governed graph does not solve the entire agent problem—nothing does—but it gives agents a coherent surface over systems that were not designed with agents in mind.</p>



<h2 class="wp-block-heading"><a href="#what-built-the-graph-first-looks-like-in-production" aria-hidden="true" class="aal_anchor" id="what-built-the-graph-first-looks-like-in-production"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>What &#8220;built the graph first&#8221; looks like in production</h2>



<p class="wp-block-paragraph">American Airlines is a useful example because its story does not begin with an LLM, but it ends with one.</p>



<p class="wp-block-paragraph">Its core system of record is SABRE, which dates to the 1960s. Years before every software roadmap acquired an “agentic” section, the company built a federated graph spanning customer, loyalty, reservations, and check-in experiences across web, mobile, and kiosks. When its CEO later asked the organization to put AI agents into production on a 90-day delivery cycle, the company did not also have to invent a context layer from scratch. It had already done the foundational architecture work.</p>



<p class="wp-block-paragraph">As Senior Principal Architect Pragadeesh Prakasam recently explained at the <a href="https://www.apollographql.com/events/the-api-layer-is-the-agent-layer-whats-standing-between-your-ai-agents-and-produc">2026 AI Agent Conference</a>, the team placed <a href="https://www.apollographql.com/apollo-mcp-server">Apollo’s MCP Server</a> on top of the graph it already operated. <a href="https://www.apollographql.com/docs/graphos/platform/security/persisted-queries">Persisted queries</a> became the agents’ tool surface. Existing field-level security scopes applied to agent interactions. Authentication, governance, and observability were not exciting new items on an AI infrastructure backlog; they were capabilities the company already had.</p>



<blockquote class="wp-block-quote is-layout-flow wp-block-quote-is-layout-flow">
<p class="wp-block-paragraph">“We could just focus on the experience,” Prakasam said. “Not on rebuilding the plumbing.”</p>
</blockquote>



<p class="wp-block-paragraph">This is the part we think gets lost when people compare adoption curves. GraphQL did not need to arrive wearing an AI costume to be useful for AI. American Airlines built a graph to solve one generation of software problems and discovered that the same investment helped with the next one. <a href="http://booking.com">Booking.com</a>, <a href="https://www.apollographql.com/customers/how-indeed-built-an-ai-hiring-assistant-in-9-months-with-apollo">Indeed</a>, and <a href="https://www.apollographql.com/newsroom/press-releases/apollo-expands-platform-to-power-the-agentic-future-at-graphql-summit-2025">Yahoo!</a> have followed a similar pattern: build a mature, governed graph, then give agents access to that foundation.</p>



<p class="wp-block-paragraph">So yes, Gartner is right that GraphQL has not spread through agent use cases with the speed or visibility of MCP. But that is a claim about adoption speed, not about value. GraphQL can still be valuable infrastructure for agents even if agents are not yet driving much new adoption of it. In practice, some of the companies best positioned to ship agents today are the ones that built the graph first. We do not think that is a coincidence.</p>



<h2 class="wp-block-heading"><a href="#where-ai-gateways-fit-and-where-they-fall-short" aria-hidden="true" class="aal_anchor" id="where-ai-gateways-fit-and-where-they-fall-short"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Where AI gateways fit and where they fall short</h2>



<p class="wp-block-paragraph">The report’s treatment of AI gateways makes the connection even more interesting. Gartner defines an AI gateway as:</p>



<blockquote class="wp-block-quote is-layout-flow wp-block-quote-is-layout-flow">
<p class="wp-block-paragraph">“A tool that acts as an intermediary between applications and various artificial intelligence services or models. Its purpose is to simplify and manage connectivity between AI applications, agents, LLMs and enterprise applications by providing a central point to enable security, governance and observability of AI workloads.”</p>
</blockquote>



<p class="wp-block-paragraph">Most AI gateways today focus on traffic management: model routing, rate limits, token controls, and defenses against attacks such as prompt injection. We at Apollo believe that all of that is necessary. It is also insufficient.</p>



<p class="wp-block-paragraph">A gateway can control where a request goes without understanding what the underlying data means. It can know that an agent is allowed to call an API without knowing whether the agent should see a particular customer field returned by that API. Without a semantic model underneath it, an AI gateway risks becoming a very sophisticated traffic cop. That’s useful, but lacking the semantics to answer many of the authorization and governance questions that make managing enterprise agents difficult.</p>



<p class="wp-block-paragraph">Our bet is that the category will move toward combining both layers: gateway controls for AI traffic and a governed model of enterprise data and capabilities. Agents need routing and rate limits, but they also need context. They need to know what systems contain, how concepts relate, what actions are available, and what a particular user or agent is authorized to see.</p>



<p class="wp-block-paragraph">In our view, Apollo GraphOS can play that role when it is implemented as a <a href="https://www.apollographql.com/context-graph">context layer</a> for agents. It provides a centralized way to control, observe, and govern what agents do across tools and APIs, using the same federated graph that companies such as American Airlines, Indeed, and Yahoo! already operate in production. The field-level access controls, authentication, and governance are not decorations added after the agent demo gets executive attention. They are part of the graph. The ceiling is much higher with these capabilities in hand.</p>



<p class="wp-block-paragraph">MCP and GraphQL solve different and complementary problems, and American Airlines’ architecture is interesting precisely because it uses both. The hard part of enterprise agents is providing the model useful context while preserving the security, governance, and observability the organization already requires. Companies with a mature graph have a head start because they have already spent years doing that work.</p>



<p class="wp-block-paragraph">Call it the trough if you want. We call it a head start.</p>



<p class="wp-block-paragraph">If you are sorting through what GraphQL, MCP, AI gateways, and the growing pile of agent protocols mean for your architecture, Apollo is offering a complimentary copy of the full <a href="https://www.apollographql.com/resources/gartner-hype-cycle-for-apis-2026">2026 Gartner Hype Cycle for APIs</a> research. If you want to talk through how GraphQL fits into your API strategy, <a href="https://www.apollographql.com/contact-sales">get in touch.</a></p>



<p class="wp-block-paragraph"><em>Gartner, Hype Cycle for APIs, 2026, By Shameen Pillai, Mark O&#8217;Neill, 22 May 2026.&nbsp;</em></p>



<p class="wp-block-paragraph"><em>Gartner and Hype Cycle are trademarks of Gartner, Inc. and/or its affiliates.&nbsp;</em></p>



<p class="wp-block-paragraph"><em>Gartner does not endorse any company, vendor, product or service depicted in its publications, and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner publications consist of the opinions of Gartner’s business and technology insights organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this publication, including any warranties of merchantability or fitness for a particular purpose.</em></p>


What We Think the 2026 Gartner Hype Cycle Gets Right About GraphQL


<p class="wp-block-paragraph">Agent Skills reduced our AI agent&#8217;s token consumption from 65,000 to 24,000 tokens and completion time from five minutes to under two — for the exact same task with the exact same agent. Here&#8217;s what happened, why it matters, and what it means for how you build with AI agents.</p>



<p class="wp-block-paragraph">There&#8217;s a recurring question in the AI community and engineering circles right now: with frontier models becoming more capable every quarter, do AI agents actually need curated Skills and instructions? Or will raw model intelligence eventually make them unnecessary?</p>



<p class="wp-block-paragraph">We decided to test it. We ran a controlled comparison and the data speaks for itself.</p>



<h2 class="wp-block-heading" id="the-experiment"><a href="#the-experiment" aria-hidden="true" class="aal_anchor" id="the-experiment"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>The experiment</h2>



<p class="wp-block-paragraph"><a href="https://www.apollographql.com/docs/apollo-mcp-server/">Apollo MCP Server</a>&nbsp;gives AI agents a secure way to access any GraphQL API. Instead of writing custom integration code, agents explore the schema, build valid operations, and fetch data through a single MCP interface. It handles schema discovery efficiently so agents don&#8217;t need the entire schema in context, just the parts relevant to the current task.</p>



<p class="wp-block-paragraph">We used this as our test case: ask Claude Code to set up an Apollo MCP Server connected to a public GraphQL API, then use it to fetch country data. Same model, same starting conditions, same goal. The only variable was whether the agent had access to&nbsp;<a href="https://www.apollographql.com/blog/apollo-skills-teaching-ai-agents-how-to-use-apollo-and-graphql">Apollo Skills</a>.</p>



<p class="wp-block-paragraph">First, here&#8217;s the agent working without Skills:</p>



<figure class="wp-block-video"><video height="638" style="aspect-ratio: 1024 / 638;" width="1024" controls src="https://wp.apollographql.com/wp-content/uploads/2026/04/mcp-without-skills.mp4"></video></figure>



<p class="wp-block-paragraph">And here&#8217;s the same task with Apollo Skills installed:</p>



<figure class="wp-block-video"><video height="638" style="aspect-ratio: 1024 / 638;" width="1024" controls src="https://wp.apollographql.com/wp-content/uploads/2026/04/mcp-with-skills.mp4"></video></figure>



<p class="wp-block-paragraph">The results:</p>



<figure class="wp-block-table is-style-regular"><table class="has-fixed-layout"><tbody><tr><td class="has-text-align-left" data-align="left"></td><td class="has-text-align-left" data-align="left"><strong>Without Skills</strong></td><td class="has-text-align-left" data-align="left"><strong>With Skills</strong></td></tr><tr><td><strong>Time to completion</strong></td><td>5+ minutes</td><td>&lt; 2 minutes</td></tr><tr><td><strong>Token consumption</strong></td><td>~65,000</td><td>~24,000</td></tr></tbody></table></figure>



<p class="wp-block-paragraph">Both runs produced the same correct result. The difference was the path to get there.</p>



<h2 class="wp-block-heading" id="why-the-gap-exists"><a href="#why-the-gap-exists" aria-hidden="true" class="aal_anchor" id="why-the-gap-exists"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Why the gap exists</h2>



<p class="wp-block-paragraph">Watch the two recordings side by side and the pattern is obvious. Without Skills, the agent spends most of its time&nbsp;<em>researching</em>: searching the web, fetching documentation, reading through pages of results, extracting relevant pieces, then attempting an approach. Often incorrectly on the first try. It backtracks, adjusts, tries again. It gets there, but the route is wasteful.</p>



<p class="wp-block-paragraph">Think of it as off-roading versus a highway. Without Skills, many paths look plausible, but some lead to dead ends, outdated patterns, or unnecessary detours. Each wrong turn burns tokens, and the agent has no way to know the path was wrong until it gets there.</p>



<p class="wp-block-paragraph">Skills pave the road. The agent knows what knowledge is available and when to load it. No researching. No guessing. No trial and error. When the right Skill activates, the agent reads it, understands the correct approach, and executes. Same destination, far less fuel.</p>



<p class="wp-block-paragraph">That 65k-to-24k drop is a 63% reduction for the same outcome. For teams running agents at scale, that translates directly to cost savings and faster iteration cycles.</p>



<p class="wp-block-paragraph">This wasn&#8217;t a one-off. We&#8217;ve been running comparisons for months since we launched&nbsp;<a href="https://skills.sh/apollographql/skills">Apollo Skills</a>, and a few patterns keep showing up.</p>



<h2 class="wp-block-heading" id="skill-quality-matters-more-than-skill-quantity"><a href="#skill-quality-matters-more-than-quantity" aria-hidden="true" class="aal_anchor" id="skill-quality-matters-more-than-quantity"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Skill quality matters more than quantity</h2>



<p class="wp-block-paragraph">Bad Skills can do more harm than good. A poorly written Skill introduces noise into the agent&#8217;s context window (wrong patterns, outdated syntax, misleading instructions) and the agent follows them dutifully. You end up worse off than if you&#8217;d given it nothing.</p>



<p class="wp-block-paragraph">This matters even more when you consider what models already &#8220;know.&#8221; The Apollo Client knowledge baked into current models is often outdated or incorrect. Models are confidently wrong without correction. A well-maintained Skill is an authoritative, up-to-date source of truth that takes precedence over whatever the model learned during pre-training.</p>



<figure class="wp-block-image size-large"><img loading="lazy" decoding="async" width="889" height="1024" src="https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.55.33-889x1024.png" alt="" class="wp-image-25357" srcset="https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.55.33-889x1024.png 889w, https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.55.33-260x300.png 260w, https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.55.33-768x885.png 768w, https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.55.33-1333x1536.png 1333w, https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.55.33-1200x1383.png 1200w, https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.55.33.png 1482w" sizes="auto, (max-width: 889px) 100vw, 889px" /></figure>



<p class="wp-block-paragraph">The download numbers on&nbsp;<a href="https://skills.sh/">skills.sh</a>&nbsp;tell this story clearly. Our&nbsp;<a href="https://skills.sh/apollographql/skills/rust-best-practices"><code class="" data-line="">rust-best-practices</code></a>&nbsp;Skill has 5,000+ downloads, the most-downloaded Rust Skill on the platform, despite being published after several alternatives. The earlier ones were low quality. Engineers tried them, got worse results, and moved on. The same pattern holds across our GraphQL Skills:&nbsp;<a href="https://skills.sh/apollographql/skills/apollo-client"><code class="" data-line="">apollo-client</code></a>&nbsp;at 1.7K,&nbsp;<a href="https://skills.sh/apollographql/skills/graphql-schema"><code class="" data-line="">graphql-schema</code></a>&nbsp;at 1.1K,&nbsp;<a href="https://skills.sh/apollographql/skills/graphql-operations"><code class="" data-line="">graphql-operations</code></a>&nbsp;at 963, and&nbsp;<a href="https://skills.sh/apollographql/skills/apollo-mcp-server"><code class="" data-line="">apollo-mcp-server</code></a>&nbsp;at 895. Quality wins.</p>



<p class="wp-block-paragraph">The implication: treat Skills like production code. Review them. Test them. Hold them to a standard.</p>



<h2 class="wp-block-heading" id="skills-need-a-maintenance-cycle"><a href="#skills-need-a-maintenance-cycle" aria-hidden="true" class="aal_anchor" id="skills-need-a-maintenance-cycle"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Skills need a maintenance cycle</h2>



<p class="wp-block-paragraph">A Skill that helps today might hurt six months from now. Models train on increasingly recent data and reason better with each generation. The gap a Skill was designed to bridge can shrink or disappear.</p>



<p class="wp-block-paragraph">Not all Skills age the same way, though. Some compensate for things a model gets wrong today, like outdated API patterns or incorrect library usage. These have the shortest lifespan because a future model may handle them natively.</p>



<p class="wp-block-paragraph">Others encode team conventions (coding standards, naming patterns, response formatting) that no model will learn from public training data. These last longer but still need review as your own standards evolve.</p>



<p class="wp-block-paragraph">The point is to treat Skills as something you maintain, not something you write once and forget:</p>



<ul class="wp-block-list">
<li><strong>Test with and without.</strong>&nbsp;When a new model drops, re-run your tasks with the Skill disabled. If the model produces the same quality&nbsp;<em>at comparable cost and speed</em>, the Skill has done its job. Consider retiring it.</li>



<li><strong>Update when the product changes.</strong>&nbsp;If your API ships a new auth flow or deprecates an endpoint, the Skill needs to reflect that. Stale Skills produce stale code.</li>



<li><strong>Keep evals running.</strong>&nbsp;Even after retiring a Skill, keep validating that the model handles the underlying task correctly. This catches regressions before they reach production.</li>
</ul>



<p class="wp-block-paragraph">We&#8217;re building this into how Apollo Skills work. Each Skill targets the latest stable version of its product, and the team that owns the product owns the accuracy of the Skill.</p>



<p class="wp-block-paragraph">Those teams don&#8217;t manually watch for drift, though. AI-powered sync pipelines detect when product changes affect a Skill&#8217;s content. An LLM triages the diff, determines whether the existing guidance is now stale, and if so, generates an update as a pull request. Product teams review and approve rather than writing updates themselves.</p>



<p class="wp-block-paragraph">Skills stay in sync with the products they describe without becoming a maintenance burden. The guidance agents receive reflects the actual state of the tools, not a snapshot from months ago.</p>



<h2 class="wp-block-heading" id="skills-are-a-force-multiplier-for-constrained-environments"><a href="#skills-shine-in-constrained-environments" aria-hidden="true" class="aal_anchor" id="skills-shine-in-constrained-environments"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Skills shine in constrained environments</h2>



<p class="wp-block-paragraph">Not everyone runs the latest frontier model with unlimited tokens. Many teams use local LLMs, self-hosted models, or lower-tier API plans for cost, privacy, or compliance reasons. CI/CD pipelines add another constraint: automated agents running in CI often operate with smaller models and stricter token budgets to keep build costs predictable. These environments need more guidance, not less.</p>



<p class="wp-block-paragraph">Skills close the gap. They give a constrained agent access to the same curated knowledge that a larger model might have partially internalized through training. A smaller model with the right Skill can match a larger model without one, producing correct results faster and with fewer tokens.</p>



<p class="wp-block-paragraph">The CI case is worth calling out specifically. When an agent runs in a pipeline (generating code, reviewing pull requests, updating docs) every minute it spends researching is a minute a developer waits for feedback. If the agent takes five minutes instead of two, that delay compounds across every PR, every build, every developer on the team. Skills cut that wait time. Faster agents mean faster feedback loops.</p>



<p class="wp-block-paragraph">Whether it&#8217;s a smaller model, a token-limited CI job, or self-hosted infrastructure, Skills can be the difference between an agent that blocks your workflow and one that speeds it up.</p>



<h2 class="wp-block-heading" id="what-this-means-for-your-team"><a href="#what-this-means-for-your-team" aria-hidden="true" class="aal_anchor" id="what-this-means-for-your-team"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>What this means for your team</h2>



<p class="wp-block-paragraph">Both runs in our experiment produced correct results. The agent is capable without Skills. But are you willing to pay the cost of letting it figure things out from scratch every time?</p>



<p class="wp-block-paragraph">For a single ad-hoc task, maybe the difference doesn&#8217;t matter. For teams running agents across dozens of workflows daily, fewer tokens and less time per task adds up. The exact savings vary by Skill and model, but the direction is consistent: Skills cut tokens and time. Multiply that across an engineering org and the return becomes hard to ignore.</p>



<p class="wp-block-paragraph">There&#8217;s also the repetition problem. Without Skills, developers paste the same instructions into every new agent session: &#8220;use v4 of this API, not v3,&#8221; &#8220;follow this naming convention,&#8221; &#8220;don&#8217;t use the deprecated auth flow.&#8221; Every session starts from zero, and the developer becomes responsible for catching the agent&#8217;s mistakes.</p>



<p class="wp-block-paragraph">Skills encode that guidance once. The agent picks it up automatically, session after session, without the developer repeating themselves.</p>



<p class="wp-block-paragraph">The less capable the model or the tighter the token budget, the more Skills matter. They put the right knowledge in context so the agent doesn&#8217;t have to go find it.</p>



<figure class="wp-block-image size-large"><img loading="lazy" decoding="async" width="885" height="1024" src="https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.54.59-885x1024.png" alt="" class="wp-image-25359" srcset="https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.54.59-885x1024.png 885w, https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.54.59-259x300.png 259w, https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.54.59-768x888.png 768w, https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.54.59-1328x1536.png 1328w, https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.54.59-1200x1388.png 1200w, https://wp.apollographql.com/wp-content/uploads/2026/04/2026-04-02-at-11.54.59.png 1482w" sizes="auto, (max-width: 885px) 100vw, 885px" /></figure>



<p class="wp-block-paragraph">Apollo Skills are now marked as&nbsp;<a href="https://skills.sh/official">Official Skills</a>&nbsp;on skills.sh, meaning they&#8217;re published and maintained by the team that builds the technology. To get started:</p>



<pre class="wp-block-code"><code class="" data-line="">npx skills add apollographql/skills</code></pre>



<p class="wp-block-paragraph">More on how Apollo Skills work and what Apollo contributes:&nbsp;<a href="https://www.apollographql.com/blog/apollo-skills-teaching-ai-agents-how-to-use-apollo-and-graphql">Apollo Skills: Teaching AI agents how to use Apollo and GraphQL</a>.</p>


Your AI Agent Doesn’t Need to Research. It Needs to Know.

Apollo Skills: Teaching AI Agents How to Use Apollo and GraphQL

What’s New in GraphOS: Apollo Summer ’25 Release – AI-Ready API Orchestration with Enhanced MCP Server, Pricing Plans, and Performance Upgrades

Every Token Counts: Building Efficient AI Agents with GraphQL and Apollo MCP Server


<p class="wp-block-paragraph">After Apollo recently released the Apollo MCP Server that allows you to generate tools for any GraphQL API, I have been investigating MCP server development within Python and Typescript frameworks. It&#8217;s easy enough to build a simple tool, but I found myself feeling déjà vu as if I were building applications before GraphQL was released. Each tool I wrote started to get more complex and I quickly started seeing multiple API fetches happen in the same tool.</p>



<p class="wp-block-paragraph">For this blog post, we&#8217;re going to look more deeply at The Space Devs APIs and the community launch library MCP server that is built in Python with the Launch Library REST API. Then we&#8217;ll look at how GraphQL can simplify the development process and provide better tools for LLM interaction.</p>



<h2 class="wp-block-heading" id="Building-simple-tools:-A-tool-for-every-REST-API-endpoint"><a href="#building-simple-tools-a-tool-for-every-rest-api-endpoint" aria-hidden="true" class="aal_anchor" id="building-simple-tools-a-tool-for-every-rest-api-endpoint"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Building simple tools: A tool for every REST API endpoint</h2>



<p class="wp-block-paragraph">One pattern I tried developing with is making a simple tool for every Entity defined in my endpoints. This would cover returning a list of those entities in a &#8220;simplified&#8221; format that returns a subset of fields along with a detailed entity tool that returns all the fields for the entity. This is a commonly recommended pattern and here is an example of that from the launch library MCP server:</p>



<pre class="wp-block-prismatic-blocks"><code class="language-python" data-line="">@mcp_llv2.tool() async def launches_list(...) 
@mcp_llv2.tool() async def launches_detailed(...)</code></pre>



<p class="wp-block-paragraph">Overall this pattern works, but using REST means that each tool is additional code that has to be written and deployed to the MCP server. Even if you have an auto-generated REST API client, there still can be a lot of hand written code. For example, the launches_list tool contains code that is primarily related to transforming the shape of the data:</p>



<pre class="wp-block-prismatic-blocks"><code class="language-python" data-line="">items = []
for x in result.results:
    items.append(
        {
            &quot;id&quot;: x.id,
            &quot;net&quot;: f&quot;{x.net} presision: {x.net_precision.name}&quot;,
            &quot;name&quot;: x.name,
            &quot;slug&quot;: x.slug,
            &quot;status&quot;: x.status.abbrev,
            &quot;rocket&quot;: x.rocket.configuration.full_name,
            &quot;orbit&quot;: x.mission.orbit.abbrev,
            &quot;provider&quot;: x.launch_service_provider.name,
            &quot;location&quot;: x.pad.location.name,
            &quot;launch_designator&quot;: x.launch_designator,
            &quot;window&quot;: (
                f&quot;{x.window_start} -{x.window_end}&quot;
                if x.window_start and x.window_end
                else None
            ),
        }
    )
return yaml.dump(
    {
        &quot;count&quot;: result.count,
        &quot;next&quot;: result.next,
        &quot;previous&quot;: result.previous,
        &quot;launches&quot;: items,
    },
    allow_unicode=True,
    sort_keys=False,
)</code></pre>



<p class="wp-block-paragraph">Each new API endpoint brings in one or two additional tools and the pattern starts to have the LLM act as an API query planner based on these tools. There are two things to be cautious about with general REST API responses:</p>



<ol start="1" class="wp-block-list">
<li><strong>Context window bloat</strong> &#8211; as the LLM uses more tools and orchestrates the responses, additional fields that aren&#8217;t relevant to the initial question can begin to add unnecessary noise to the context window.</li>



<li><strong>API Orchestration Distractions for LLMs</strong> &#8211; each tool should encapsulate everything you need to give the proper context to the LLM for that piece of functionality. Having an LLM run through two or more tools for a given functionality because the API is broken apart is going to create additional processing overhead that is going to potentially confuse LLMs and cost more as additional tool calls are required.</li>
</ol>



<h2 class="wp-block-heading" id="GraphQL-Operations-as-MCP-Tools"><a href="#graphql-operations-as-mcp-tools" aria-hidden="true" class="aal_anchor" id="graphql-operations-as-mcp-tools"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>GraphQL Operations as MCP Tools</h2>



<p class="wp-block-paragraph">GraphQL initially became popular with web developers in the promise of providing an operation as the contract of data needed for a given UI functionality. Agentic experiences will also have many of the functionalities our websites have and it makes sense to map a GraphQL operation to that functionality. Defining GraphQL operations as MCP tools means:</p>



<ul class="wp-block-list">
<li>No writing tool code, just define the fields you need in the operation</li>



<li>The graph does the API orchestration, not the LLM</li>



<li>Adding or updating tools can be done without code deployments</li>
</ul>



<p class="wp-block-paragraph">To demonstrate a comparison, we created a GraphQL API for the Launch Library to create MCP tools with. In the Apollo MCP Server repository is an example operation that can be used to create an MCP tool for searching upcoming launches. GraphQL provides nice documentation capabilities that we can use to make tool descriptions. We default to comments and return type information defined in your schema but you can simply override that description by adding a comment to the top of your operation.</p>



<figure class="wp-block-image size-large"><img loading="lazy" decoding="async" width="1024" height="563" src="https://wp.apollographql.com/wp-content/uploads/2025/05/comparison-1024x563.png" alt="" class="wp-image-24348" srcset="https://wp.apollographql.com/wp-content/uploads/2025/05/comparison-1024x563.png 1024w, https://wp.apollographql.com/wp-content/uploads/2025/05/comparison-300x165.png 300w, https://wp.apollographql.com/wp-content/uploads/2025/05/comparison-768x422.png 768w, https://wp.apollographql.com/wp-content/uploads/2025/05/comparison-1536x845.png 1536w, https://wp.apollographql.com/wp-content/uploads/2025/05/comparison-1200x660.png 1200w, https://wp.apollographql.com/wp-content/uploads/2025/05/comparison.png 1920w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure>



<p class="wp-block-paragraph">A nice benefit about this is how the LLM can use guided context to help re-execute the tools we defined. For example, we provide some information around the search term being best as codes or single words. The LLM understands this tool is best for short searches and can re-run them.</p>



<figure class="wp-block-image size-full"><img loading="lazy" decoding="async" width="1000" height="1000" src="https://wp.apollographql.com/wp-content/uploads/2025/05/spacedevs.gif" alt="" class="wp-image-24350"/></figure>



<h2 class="wp-block-heading" id="Improved-developer-experience"><a href="#improved-developer-experience" aria-hidden="true" class="aal_anchor" id="improved-developer-experience"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Improved developer experience</h2>



<p class="wp-block-paragraph">Along with having no code for your MCP tools, writing those MCP tools is simple. GraphQL endpoints have options like <a href="https://thespacedevs-production.up.railway.app/">Apollo Explorer</a> that give you a sandbox to author and test out operations that you could then run as tools. If you open the <code class="" data-line="">/graphql/TheSpaceDevs</code> folder up in VS Code, you’ll see how you can have completion options for fields available in your schema as you write the operations:</p>



<figure class="wp-block-image size-large"><img loading="lazy" decoding="async" width="1024" height="678" src="https://wp.apollographql.com/wp-content/uploads/2025/05/Screenshot-2025-05-18-at-11.16.21 PM-1024x678.png" alt="" class="wp-image-24352" srcset="https://wp.apollographql.com/wp-content/uploads/2025/05/Screenshot-2025-05-18-at-11.16.21 PM-1024x678.png 1024w, https://wp.apollographql.com/wp-content/uploads/2025/05/Screenshot-2025-05-18-at-11.16.21 PM-300x198.png 300w, https://wp.apollographql.com/wp-content/uploads/2025/05/Screenshot-2025-05-18-at-11.16.21 PM-768x508.png 768w, https://wp.apollographql.com/wp-content/uploads/2025/05/Screenshot-2025-05-18-at-11.16.21 PM.png 1194w" sizes="auto, (max-width: 1024px) 100vw, 1024px" /></figure>



<h2 class="wp-block-heading" id="MCP-Tool-Generative-Capabilities"><a href="#mcp-tool-generative-capabilities" aria-hidden="true" class="aal_anchor" id="mcp-tool-generative-capabilities"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>MCP Tool Generative Capabilities</h2>



<p class="wp-block-paragraph">You might also want to just have the LLM be a little more curious and learn about the schema to generate operations for tools or even to just dynamically execute. The <code class="" data-line="">--introspection</code> option allows the LLM to enter the schema at any node and navigate the tree to keep the context window concise to the task at hand.</p>



<p class="wp-block-paragraph">Here&#8217;s an example of the tool in action with Claude:</p>



<figure class="wp-block-image size-full"><img loading="lazy" decoding="async" width="1920" height="1080" src="https://wp.apollographql.com/wp-content/uploads/2025/05/output.gif" alt="" class="wp-image-24354"/></figure>



<h2 class="wp-block-heading" id="Conclusion"><a href="#conclusion" aria-hidden="true" class="aal_anchor" id="conclusion"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Conclusion</h2>



<p class="wp-block-paragraph">When you need to write MCP tools for your APIs, GraphQL gives you a serious advantage in rapid iteration. A series of tools can be created in minutes and getting started is just a single command. You can <a href="https://www.apollographql.com/docs/apollo-mcp-server/quickstart">get started today</a> with any GraphQL API, no API key or account required.</p>



<p class="wp-block-paragraph"><a href="https://www.apollographql.com/events/get-started-with-mcp-tools-for-your-graphql-api-today">Check out</a> my latest Apollo livestream from 5/21/2025 where I dive into developer workflows like hot-reloading with your MCP server with some Q&amp;A at the end for MCP questions.</p>



<p class="wp-block-paragraph"></p>


Building MCP Tools with GraphQL: A Better Way to Connect LLMs to Your API


<p class="wp-block-paragraph">Today we&#8217;re introducing <a href="https://github.com/apollographql/apollo-mcp-server">Apollo MCP Server,</a> enabling AI systems like Claude and GPT to interact with your organization&#8217;s APIs through GraphQL. While you can learn more about how to get started here, I wanted to share why we think GraphQL enabled by Model Context Protocol (MCP) will be an essential part of a modern AI stack.</p>



<h1 class="wp-block-heading"><a href="#connecting-ais-to-apis" aria-hidden="true" class="aal_anchor" id="connecting-ais-to-apis"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Connecting AIs to APIs</h1>



<p class="wp-block-paragraph">MCP – a new protocol that standardizes how LLMs interface with external systems – opens the door wide open to agentic applications. APIs are the entry point to capabilities: adding items to carts, checking order status, scheduling appointments, updating inventory. When AI can reliably interact with these systems, we unlock an entirely more capable kind of software where natural language becomes the interface to complex business operations. MCP makes this possible by providing the connective tissue between AI&#8217;s language understanding and your API infrastructure.</p>



<p class="wp-block-paragraph">One approach to MCP is to build an MCP server for each API you want AI to access. Your payment API gets its own MCP server, your inventory system gets another, customer data gets a third. It&#8217;s a straightforward way to get started. Most teams can build an initial MCP for a service reasonably quickly and LLMs support multiple MCP tools. It makes for a compelling experience.</p>



<p class="wp-block-paragraph">As we’ve worked with teams pushing from prototype to product, we&#8217;re discovering that complex AI systems need more than what is possible by directly connecting individual APIs through MCP. We&#8217;ve found four particular requirements that show up consistently:</p>



<p class="wp-block-paragraph"><strong>Deterministic execution</strong>: An AI will often have to call multiple APIs in sequence and filter the results to just the appropriate fields for the user. This execution pattern needs to work exactly the same way every time for consistent, correct user experiences. Consider an AI returning purchase history – you can&#8217;t have one customer seeing five past orders while another sees three, or some results including shipping details while others don&#8217;t. We&#8217;ve also found that the more focused you keep the AI&#8217;s context, the better results you get. When APIs return large objects with hundreds of fields, the extra data can lead AI down unintended paths. Limiting data to exactly what&#8217;s needed keeps the AI on track and improves overall response quality.</p>



<p class="wp-block-paragraph"><strong>Policy enforcement</strong>: Giving AI direct access to individual APIs makes it difficult to enforce appropriate limits and access controls. A rule like “only loyalty members have access to the detailed shipping tracking information for their outstanding orders” can’t easily be written in the API layer, as it relates to a specific combination of data spread across multiple systems.</p>



<p class="wp-block-paragraph"><strong>Efficiency</strong>: Even simple customer queries can generate thousands of tokens when APIs return complete objects. Processing all of this and directly orchestrating multiple tool invocations from the AI becomes costly, both in terms of tokens and milliseconds.&nbsp;</p>



<p class="wp-block-paragraph"><strong>Agility</strong>: The pace of AI is unprecedented. The technology and techniques are progressing furiously. Just as interesting to me is that the freeform nature of AI makes it possible to deliver new functionality faster than ever before: most users wouldn’t tolerate a mobile app that changed every day, but continuous improvement of an AI agent is nothing but good news. So an AI stack needs to accommodate and facilitate a tempo we’ve never seen before in software development. This requirement rules out strategies that aren’t aligned to the modern platform engineering discipline of self-service tools and workflows.</p>



<h1 class="wp-block-heading"><a href="#graphql-solves-these-needs-well" aria-hidden="true" class="aal_anchor" id="graphql-solves-these-needs-well"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>GraphQL solves these needs well</h1>



<p class="wp-block-paragraph">What these requirements point to is the need for an abstraction layer between the AI and APIs. This <a href="https://www.apollographql.com/guides/api-orchestration">orchestration layer</a> exposes a set of MCP tools, where each tool provides a specific capability backed by some number of orchestrated API calls. The AI&#8217;s job is to reason about which MCP tools to invoke; the orchestration layer’s job is deterministic, efficient execution and policy enforcement.</p>



<p class="wp-block-paragraph">GraphQL is an ideal foundation of this orchestration layer.</p>



<ul class="wp-block-list">
<li>It’s declarative: MCP tools are defined in terms of schema and queries, rather than code.</li>



<li>It’s performant: one GraphQL query can traverse your entire graph, gathering exactly what&#8217;s needed (and nothing that’s not) in an optimized way.</li>



<li>It’s self-documenting: LLMs can reason about the meaning of the graph and its objects.</li>



<li>It&#8217;s standards-based: a mature ecosystem with proven scalability.</li>



<li>It’s self-service: on a GraphQL-based platform like Apollo, everyone with access to the graph can wire up MCP tools with the appropriate level of governance.</li>
</ul>



<p class="wp-block-paragraph">The characteristics of GraphQL align perfectly with what AI needs to access APIs: a clean, understandable interface to AI while handling the messy reality of existing systems and services. </p>



<p class="wp-block-paragraph">Consider the earlier example of retrieving purchase history with shipping updates. A GraphQL query defines an MCP tool that the AI system can use. Under the hood, that query orchestrates multiple API calls into the order, shipping and loyalty systems and enforces the necessary policy check. That query can be authored and deployed by anyone on the platform, with the support of design and governance tools provided by the graph platform.</p>



<p class="wp-block-paragraph">The challenges an abstraction layer solves may seem like tomorrow’s problems when just getting started with MCP. It’s always worth planning for success, but a good declarative architecture also saves time and effort in the earliest stages of development. A GraphQL-based approach avoids having to build, debug and deploy custom MCP servers. Apollo Connectors let you bring REST APIs to GraphQL (and now MCP) without any modification. And the benefits of having a semantic API layer are truly remarkable – watching an LLM traverse a large GraphQL schema as it builds a new query is an exciting, eye-opening moment.</p>



<h1 class="wp-block-heading"><a href="#introducing-apollo-mcp-server" aria-hidden="true" class="aal_anchor" id="introducing-apollo-mcp-server"><svg aria-hidden="true" class="aal_svg" height="16" version="1.1" viewBox="0 0 16 16" width="16"><path fill-rule="evenodd" d="M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z"></path></svg></a>Introducing Apollo MCP Server</h1>



<p class="wp-block-paragraph">The Apollo MCP Server represents our continued investment in making GraphQL the best way to build modern API infrastructure. We&#8217;ve spent years developing tools that help teams build, manage, and scale GraphQL – from <a href="https://www.apollographql.com/docs/react">Apollo Client</a> to <a href="https://www.apollographql.com/federation">Apollo Federation</a> to our most recent innovation, <a href="https://www.apollographql.com/graphos/apollo-connectors">Apollo Connectors</a>. Now we&#8217;re extending that toolkit to address AI integration.</p>



<p class="wp-block-paragraph">By implementing MCP for GraphQL, we&#8217;re making it simple to connect AI to your existing graph. If you&#8217;re already using Apollo, this fits naturally into your stack. If you&#8217;re new to GraphQL, the MCP server provides a compelling entry point that solves immediate AI challenges while setting you up for future growth, all on top of the APIs you already have.</p>



<p class="wp-block-paragraph">We believe the architectural patterns emerging around AI – the need for precision, policy, efficiency, and agility – align perfectly with what GraphQL was designed to do. This MCP server is our way of helping teams realize those benefits.</p>



<p class="wp-block-paragraph">As you explore connecting AI to your services, consider not just immediate integration but your longer-term strategy. The architectural decisions you make today will determine how quickly and safely you can evolve AI capabilities tomorrow.</p>



<figure class="wp-block-table"><table class="has-fixed-layout"><tbody><tr><td>Ready to get started? Connect your REST and GraphQL APIs to the <a href="https://github.com/apollographql/apollo-mcp-server">Apollo MCP server </a>today, explore our Odyssey training course, or join us at our upcoming <a href="https://www.apollographql.com/events/get-started-with-mcp-tools-for-your-graphql-api-today">event</a> to learn more.&nbsp;<br><br>The future of APIs is here – powered by <a href="https://www.apollographql.com/guides/api-orchestration">graph-based API orchestration</a> that makes it easier than ever to build applications.</td></tr></tbody></table></figure>



<p class="wp-block-paragraph">​​</p>


Artificial Intelligence

Apollo Summit 2026: Turn Your API Platform Into Your AI Platform

What We Think the 2026 Gartner Hype Cycle Gets Right About GraphQL

Your AI Agent Doesn’t Need to Research. It Needs to Know.

Apollo Skills: Teaching AI Agents How to Use Apollo and GraphQL

What’s New in GraphOS: Apollo Summer ’25 Release – AI-Ready API Orchestration with Enhanced MCP Server, Pricing Plans, and Performance Upgrades

Every Token Counts: Building Efficient AI Agents with GraphQL and Apollo MCP Server

Building MCP Tools with GraphQL: A Better Way to Connect LLMs to Your API

Past posts

The Future of MCP is GraphQL

Getting Started with Apollo MCP Server for any GraphQL API

Artificial Intelligence

Apollo Summit 2026: Turn Your API Platform Into Your AI Platform

What We Think the 2026 Gartner Hype Cycle Gets Right About GraphQL

Your AI Agent Doesn’t Need to Research. It Needs to Know.

Apollo Skills: Teaching AI Agents How to Use Apollo and GraphQL

What’s New in GraphOS: Apollo Summer ’25 Release – AI-Ready API Orchestration with Enhanced MCP Server, Pricing Plans, and Performance Upgrades

Every Token Counts: Building Efficient AI Agents with GraphQL and Apollo MCP Server

Building MCP Tools with GraphQL: A Better Way to Connect LLMs to Your API

Past posts

The Future of MCP is GraphQL

Getting Started with Apollo MCP Server for any GraphQL API

Stay in our orbit