800 MCP servers and 5,000 tools behind one internal gateway. That's the number Uber published this week for its own agent infrastructure, and it says something sharper than another "we adopted MCP" post: tool sprawl inside a single company can now rival the size of a public API ecosystem. The mechanism is unglamorous, and that's the point. A registry acts as the control plane, tracking ownership and which tools are enabled for which agents. A proxy gateway sits in the data plane, translating MCP calls into whatever the backend actually speaks, HTTP, gRPC, TChannel, routed through Uber's existing Muttley service mesh. Authorization happens per tool, not per service, and PII gets redacted from responses before an agent ever sees it. An AutoCrawler component pulls API definitions straight from Uber's IDL registries so teams don't have to hand write MCP wrappers for things that already exist. Uber isn't alone in deciding MCP needed a front door. Cloudflare published its own enterprise reference architecture the same season, built around a different bet: distribute MCP servers globally instead of centralizing them, and fix the token bill before the security bill. Its Code Mode approach claims a 94% cut in context consumption, from roughly 9,400 tokens down to 600 for the same 52 tools, by making agents discover tools progressively instead of dumping every schema up front. Speakeasy, MintMCP and a few others are now publishing straight up buyer's guides for "best MCP gateway," which is usually the tell that a pattern has become a market. What nobody's pricing in yet is the irony sitting underneath all of it. MCP got popular because it let any team stand up a tool in an afternoon without asking platform engineering for anything. The gateway is the bill for that freedom: a whole second system whose entire job is reintroducing the governance everyone skipped the first time. My read: this is where agent infrastructure spending actually goes in year two. Not more model calls, the plumbing that makes five thousand tools safe to point an agent at. x.lingyaoai.com/UberEng/status/2106072…

Private drinks with the private network. Join the @Speakeasydev and @Tailscale teams at a real life speakeasy. It's on the last day of @aiDotEngineer, we'll be chatting about privacy and security in the age of agents. Space is limited: luma.com/po85yuqq

🎯You are so right. There's so much intelligence wrapped up in those chats. We capture all sessions across Anthropic, OpenAI Cursor and other providers and tag cost, token use tool calls and various security policy evaluations. @ghaidar0

Skills are a powerful primitive, but organizations have struggled to get a handle on them. Extending MCP to cover Skills is going to encourage more companies to manage skills as a 1st class artifact, and that's going to explode usage. With MCP there was enough friction in creating them that most employees wouldn't create their own. Not the case with skills, anyone can create a markdown file, so management has been highly dispersed. Not a single company we've worked with knew how many skills were in active use across their org (before implementing @speakeasydev ofc). That's not inherently a problem, but there's usually a lot of duplicate work being done. Once companies start bundling skills in their MCP servers they'll start spending more time curating them. Usage will go up without users even needing to be aware. Kudos to the committee for continuing to be great stewards of the protocol.

This is an exciting update for claw devotees who want to pitch this upwards to their IT and Security leaders. There's still a major gap though. Claws are long lived and proactive entities that will work and take action like employees. For humans you can go into @okta and authorise access to apps. How do you do this for long lived agents like openclaw? We've been cooking on an easy to use Agent Identity feature on the @speakeasydev control plane. I use the word "easy" because the Agent Identity space is filled with dozens of acronyms: Workloads, SPIFFE, WHIMSE, CXA, ID-JAG and more. We've made built on the state of the art protocols here to give users a simple workflow to register an agent (or discover it automatically) and grant it scoped and revocable permissions to various mcp servers and other tools. Every agent has its own identity and short live tokens rather than less secure shared access models like service accounts. For an IT admin though its just a few clicks or running a skill. I'm excited to see how many enterprises we can accelerate agent adoption in by solving the foundational building block of identity. A little more on how we've done that for Claude Tag. Open Claw next! 🦞speakeasy.com/blog/agent-ide…

SaaS as a category was always silly. It bundled systems of record like Slack or Jira with productivity apps like Miro or Canva. Agents have made this division clear... Systems of record are more important than ever, just look at MCP usage on the @speakeasydev gateway Productivity apps, not so much...

I understand the rush for companies to adopt token-based pricing, but I can't help thinking that for most it will turn out to be a mistake in the long run for the simple reason that, 1 Claude token ≠ 1 DeepSeek token Different models use tokens in completely different ways. That doesn't matter if you build your product on a single model, but what happens when you change the underlying model? Are you going to switch all your billing? What if your product becomes multi-model? What if you allow users to BYOM? It gets complicated quickly. For most products, token-based pricing doesn't make sense because tokens are not tied to the economic value of the outcome the product is driving. I saw recently that @PostHog's self-driving feature charges $15/PR opened i.e. patches submitted. I think this is much more appropriate for the majority of AI-based products. Focus on the outcomes you are driving and price-in what it costs for you to accomplish it. The one place where token-based pricing does make sense is companies where the stream of tokens is the basis of the value. For example a model router company, or a company like @Speakeasydev where the greater % of the companies AI usage flowing through the platform, the more value the customer gets.

Nvidia pushes agent containment below the model, dv01 and Cognizant put agents into finance and claims workflows, and Google tests buying from Flipkart inside Gemini. Speakeasy adds tighter identity and policy controls for MCP.

@MoonPay distributed AI to every employee without turning security into the bottleneck. Now they run 200+ MCP servers and 60,000+ agent sessions through one control plane. Great writeup from @AgenticAIFdn:

Most of the @speakeasydev has been together for multiple years now through cycles of growth and PMF iteration. Longevity >> team size.

We'll be at The @aiconference next week at Pier 48, San Francisco. Come find us on the floor. We'll be talking through how teams put identity, policy, and an audit trail on every agent action. We also have a shuffle board. See you there!

Promise its costs not our revenue 😂

Very happy we partnered with @speakeasydev on this, previous SDK vendor risk was very real for the ecosystem!

Open tooling builds stronger developer ecosystems. We teamed up with @speakeasydev to ship the new Google GenAI SDKs for our Interactions, Agents, and Webhooks APIs. Read why we believe SDK generation belongs in the open and explore Speakeasy's newly open-sourced OpenAPI generator suite: x.lingyaoai.com/GoogleAIStudio/status/…

Working with the @GoogleAIStudio team has been a dream. We love working with companies that care about DX & AX as much as we do. We're just getting started, and can't wait to keep building alongside them!

Excited to share that we partnered with @speakeasydev to build the sdks for the Interactions API and supported them to open-source their their entire OpenAPI generator suite (SDKs, agent CLIs, and MCP servers) open source for the community to use: developers.googleblog.com/wh…

Today we're announcing our partnership with @googledevs, and open sourcing infrastructure for agent experience together 🤝 Every enterprise is coming to the realization that developer experience is no longer the whole story. Agents are rapidly becoming the users integrating with your systems, and they need infrastructure built for them: Discoverable interfaces, structured output, context they can query at runtime. Those are the foundations we've been building together. Google is future-proofing the Gemini APIs' integration experience so agents get the same first-class treatment developers do, and Speakeasy is the platform generating and governing the artifacts that make that possible. We've been building the Speakeasy AI Control Plane to provide enterprises the identity, policy enforcement, and observability over every agent connection. But before you implement control, you need robust connections. We've been working with the Google team to Open Source the primitives needed for agents to build those connections. The Speakeasy generator is now open source and includes everything you need to make your services consumable by agents: MCPs, CLIs, and SDKs speakeasy.com/blog/partnerin… github.com/speakeasy-api/ope…