Arcade.dev | The MCP Runtime

The thin line between magic and tragic: getting MCPs right. Tomorrow at 4:20 PM @TheMostlyGreat is moderating this panel at the AI Agent Conference in NYC with teams from Groupon, PagerDuty, and HubSpot. 📍Trianon Ballroom.

"An MCP server is not compliant just because it connects." That's how @sterlingdreyer opened his talk at the first MCP Dev Summit — with the spec authors sitting in the room. Then he showed the data from scanning 40,000+ MCP servers. 🧵👇

Last night in Plano, we hosted a dinner w/ AI leaders from Morgan Stanley, Walmart, Gartner & more. Key themes: governance first, agent memory at scale, and closing the gap between what agents can do & what enterprises can safely let them do. More of these coming 🚀

Our CEO @TheMostlyGreat is joining @Redisinc, FastMCP, @AgenticAIFdn and @wandb by @CoreWeave on May 5th for a panel on MCP vs. A2A. We’ll cover how these approaches show up in production and what teams should understand before committing to one. redis.io/resources/vide…

Tavily is now integrated into @TryArcade , which will enable secure, production-ready workflows. From financial research and CRM enrichment to healthcare, legal, and market intelligence, teams can automate complex, multi-step workflows with real-time data, secure execution, and fully auditable actions. Check out a full rundown here: tavily.com/blog/tavily-is…

Good @WSJ piece by @mims on AI reliability. The models are ready. The tooling is ready. The layer most enterprises are still missing: who can this agent act for, on which systems, with what permissions? Solve that and the gains are real. wsj.com/tech/ai/ai-mod…

"The model is the brain. The harness is the body." I sat down with @robinnewhouse from @cline We got into harness design, model-agnostic agents, and why the bottleneck in coding agents has quietly shifted. 🧵👇

Most AI SRE tools fail because they try to be "autonomous." I am skeptical of that. A model can't remediate production unsupervised, and it shouldn't. The real win is the "Companion" model. The AI does the legwork (triage, timelines, correlation). The human makes the judgment. I broke down 5 #AI #SRE workflows and the #MCP runtime infrastructure needed to make them safe in production on the Arcade blog. The goal: let the on-call start on page 5 instead of page 1. Full post here: arcade.dev/blog/claude-co… #AI #SRE #RELIABILITY #ONCALL #DEVOPS #MCP

Next week at the AI Agent Conference, @TheMostlyGreat is moderating a panel on running MCP in production at scale. Groupon. PagerDuty. HubSpot. Teams that have actually shipped. Tuesday, 5/5, 4:20 PM, Trianon Ballroom.

Why does MCP work in local testing, but break the moment a real client connects over HTTP with authorization involved? That gap between “it should work” and “it actually works in production” is where most MCP debugging challenges show up. Read the blog by Nate Barbettini Founding Engineer of @TryArcade: bit.ly/4mN8fhX

Yesterday, the Arcade team rang the @Nasdaq Closing Bell. Every enterprise is figuring out agents right now. We've been building for this moment since day one. Thank you @Wing_VC and @EricNewcomer.

Anthropic just published how they built Managed Agents. They solved the orchestration layer elegantly. The hands are still an exercise for the reader. @TheMostlyGreat wrote about what that means: arcade.dev/blog/anthropic…

The gap between a working demo and production is where most AI initiatives die. 📍April 29, Plano, TX. Arcade.dev, @LangChain_OSS, and @Redisinc are hosting a small, invite-only dinner for senior AI leaders working through that. Apply: luma.com/sm64r4mg

The Vercel breach isn't an AI incident. No model jailbroken. No prompt injection. No MCP exploit. - Context.ai had broad Workspace scopes + long-lived refresh tokens in a shared AWS env - Attackers grabbed a Vercel employee's token - Pivoted into Vercel env vars

As long as you own the context, you own the agent. Join me in today's webinar to learn how to implement this! 📺 Live at @TryArcade's YouTube channel at 13:30 PST

Full breakdown from @TheMostlyGreat, including why most "skills" stories you're hearing right now don't survive contact with reality: youtu.be/-GowfW_T16U

MCP gateway = necessary. Not sufficient. What you actually need is a control plane. That's a runtime, not a proxy.

Hot take from the first-ever MCP Dev Summit in NYC: The MCP gateway is not the control plane. 🧵