Inside Freshworks' MCP Gateway

From MCP Discovery To Execution: Building a Governed Marketplace & Gateway for Agentic Systems - Rahul Ganesh Partheeban, Freshworks MCP is a protocol, not a platform. In this talk, Rahul Partheeban from Freshworks shares the real story of taking Model Context Protocol from a LinkedIn doom-scroll discovery in early 2025 to a production enterprise deployment serving more than 75,000 global customers. Rahul works in platform engineering on marketplace and integrations, and this is a hands-on look at the patterns enterprises actually need around MCP. The premise: human support and service agents work across Stripe, Shopify, ShipRocket, Datadog, PagerDuty and dozens of other tools every day. When you move to AI agents, they need that same access at scale, and the raw protocol does not give you authentication, routing, governance, or trust. Freshworks solved it with two layers: a governed marketplace for discovery and a gateway control plane for execution. What Rahul covers: Protocol vs platform: Why the MCP spec is a guide for agent-to-tool communication, not an answer to the problems you hit in production. The journey: Tracking the spec from local-only to remote, SSE, and the March 2025 streamable HTTP launch that made serverless, enterprise-scale MCP possible. Apps as MCP tools: Reusing 15 years of battle-tested marketplace, OAuth, and remote functions instead of reinventing the wheel. The gateway pattern: A control plane, not a proxy, providing a unified MCP endpoint, multi-tenant routing, session context, rate limits, retries, and timeouts. Distributed session store: Keeping agentic systems stateful and horizontally scalable across pods, restarts, and rolling deploys even as the protocol moves toward stateless. Remote servers as black boxes: Why circuit breakers, the noisy-neighbor problem, and timeouts matter when you do not control third-party code. Three-tier error taxonomy: Separating protocol, gateway, and application errors so agents can intelligently retry and recover workflows. Discovery and the MCP registry: Why the official registry solves finding servers but not trusting them. A governed marketplace: Per-tenant vetting, tenant-level secret isolation, OAuth 2.1, and DCR for one-click install. Why DCR matters: How dynamic client registration removes the painful manual OAuth credential setup and lifts the adoption curve. If you are building enterprise AI agents, running an MCP platform, or wrestling with governance, trust, and production reliability, this talk lays out a complete tool lifecycle: marketplace for discovery and trust, gateway for execution and reliability. Links & Resources: Freshworks: https://www.freshworks.com Model Context Protocol: https://modelcontextprotocol.io MCP Registry: https://registry.modelcontextprotocol.io MCP Specification: https://spec.modelcontextprotocol.io OAuth 2.1 Dynamic Client Registration (RFC 7591): https://datatracker.ietf.org/doc/html... Timestamps: 00:00 Intro - finding MCP on LinkedIn 01:00 The premise: MCP is a protocol, not a platform 02:00 How human agents work across business tools 02:45 The journey: tracking the MCP spec in early 2025 03:25 Streamable HTTP and scaling in serverless 04:45 Exposing apps as MCP tools first 05:55 The gateway as a central control plane 07:00 Why a platform handles identity and OAuth 08:30 Why you need a gateway between agent and servers 09:55 The distributed session store 11:50 Remote servers as black boxes: circuit breakers 13:00 Rate limits, timeouts, and analytics 13:40 Three-tier error taxonomy 15:20 Discovery and the MCP registry 16:40 Building a governed marketplace 18:00 OAuth 2.1, DCR, and one-click install 19:55 The full tool lifecycle and why it matters 21:15 Closing and questions #ModelContextProtocol #AIAgents #EnterpriseAI