Ravi Madabhushi @ravibits

This'll be my third @mcpsummit! Every time, the MCP conversation has moved somewhere new and I leave with more to think about. On June 15th I'm running a session on MCP as it actually is today: the wins, the messy bits, where it goes next. If you're there, come say hi 👇

we shipped a feature nobody asked for. three of our best customers said it was exactly what they needed. listen to users. but also just know things.

@sevaustinov @garrytan 10/10 relate! Writing the skill is the easy 20%. The other 80% is everything you only discover by using it in anger. I now straight away get agents to update, stress-test on subagents, and loop until green. That one change is why my skills hold up now!

I’m excited to introduce you to Von (vonlabs.ai)! Last year we decided to pivot from Rattle to Von. AI capabilities have come a long way since 2022, and no matter how much we tried to make Rattle feel AI-native, it still felt like driving a Ferrari at 20 MPH. So we decided to build a new product from the ground up, one that was built for this AI world. That product is Von. Von is the intelligence layer over the entire GTM tech stack. It is the brain, orchestration, and execution layer - all in one. It is built for RevOps teams that never have enough hours or hands to go through everything that is on their plates. The way Von is built, it can handle 80% of all of RevOps operational work - freeing them to focus on strategy and go through the backlog that is years long. RevOps being short on capacity doesn’t hurt just the RevOps teams. It hurts the business users - the CRO, CMO, VP Sales, Sales & CS Manager, and frontline teams who get stuck in the backlog. These are the leaders who have to choose what gets done and make tradeoffs on what gets analyzed. To them, Von feels like having their own VP of RevOps - one that understands the business, understands them, and is available 24/7. If you talk to any Von customer, they will tell you it felt too good to be true till they tried it. It feels like a headcount on their team, a business partner. As part of today’s launch, we have launched a new website (vonlabs.ai) detailing how it works, what it can do, and the impact it is driving for customers, amongst many other things. I hope you will take a minute out of your day to visit it. Rattle was built to help RevOps teams. Von is built to multiply the impact of RevOps 100x.

@RhysSullivan Like folks have already stated, plugins enable you to assemble skills. A single canonical reference can be fed into many skills without replicating the content.

@RajanAnandan @prukalpa @AtlanHQ "The context layer is not a feature you ship." This line should be on a wall in every AI platform team's office. Models commoditize. Context compounds. That's the whole game.

@akothari @NotionHQ @ivan That's a mystery solved. Such a cool story! 😁

5/ The fix is boring: each user authenticates their own connection, the agent acts as that person, the credential never enters the execution context. Easy to build. Painful to retrofit. An afternoon at 30 users, a proper project at 3,000. Worth doing before the bill arrives. scalekit.com/blog/admin-acc…

4/ And it surfaces at the worst time. Your first serious enterprise prospect asks who changed what, your audit log says "admin@ updated 23 records," and now you're building a permissions layer in the middle of a sales cycle.

1/ Most teams I talk to aren't bolting an agent onto their product. The agent IS the product. They ship first and fix things only when those things start to matter. Usually the right call. On one thing it isn't, and it sneaks up on people.

3/ Agents kill that guarantee. The whole point is that they decide what to do at runtime. So the credential they hold has admin access to everything, at every step of a process you can't predict, kicked off by a user who should only touch a sliver of it. That gap is the real problem.

2/ A normal Zapier flow runs the same steps in the same order every time. So even if the credential behind it can touch way more than it should, it never will. The flow only does what you configured. Broad access, bounded behaviour. That's what made "just use the admin account" safe to forget.

The question to ask any vendor handling OAuth tokens on your behalf, Scalekit included, is what happens after the token is decrypted. Who can reach it, through what path, and what's the blast radius if something goes wrong. Full architecture and checklist: bit.ly/4wU7AzW. Migrating off Composio? DM me.

OAuth breaches have become routine news. This week hit closer to home. Composio, a player in our space, had a breach. Several customers and prospects reached out with questions they hadn't asked this seriously before. The short version: encryption at rest is necessary but not sufficient.

@toddjcrane @steipete Ha, thanks. A cringe-free pitch is genuinely hard. Happy to nerd out on any of the above if it's useful!

We've already seen custom connectors go as specific as Salesforce's Metadata API. That's just the start.

An agent is only as useful as the systems it can act on. Scalekit now lets you bring any API or MCP server in as a first-class connector. Third-party or internal. You define the auth. Everything else works the same as built-ins. The connector catalog is now yours to own.

If a connector wasn't in our catalog, your team waited on our release cycle. That changes today. Custom connectors in Scalekit: your approval workflow API, your partner data systems, your self-hosted MCP server. Show up in the dashboard alongside everything else. Nothing else changes.

Agent reliability shouldn't be something you reactively chase. It should be something you can systematically improve.

6/ This is the observability gap most agent infrastructure hasn't closed yet. Today we're introducing Agent tool observability in Scalekit: tool call volumes, error rates, and failure logs with enough runtime context to act on.

As agents take on harder tasks in your real workflows, the trace you have on them starts to matter in ways it didn't six months ago. 1/ Traditional API monitoring works because the call path is fixed. Same endpoint, same parameters, reproducible failure. You can write an alert. You can reproduce it in staging.