cd ~/
Agentic AIMulti-Agent SystemsMCPPlatform ArchitectureInfrastructure

MCP Won

MCP hit 97 million monthly downloads and Salesforce just went fully agent-native. The integration layer for agents is decided.

Evyatar Bluzer
3 min read

The integration problem for AI agents is effectively solved. MCP won.

The Numbers

Model Context Protocol hit 97 million monthly SDK downloads in March, up from 2 million at launch in November 2024, which makes it one of the fastest adoption curves in open-source history. Raw downloads only get you so far as a signal, though. The number I actually watch is who ships MCP servers, and as of last week the answer is everyone.

Salesforce Just Went Headless

At TDX on April 15, Salesforce announced Headless 360. The entire platform - every data object, workflow, and business logic block - is now exposed as APIs, CLI commands, and 60+ MCP tools, so coding agents like Claude Code, Cursor, and Codex get native Salesforce access through the protocol. The framing matters here: Salesforce rebuilt the platform so agents can operate it directly, where a human used to click through a browser.

The same week, Databricks rebranded its AI Gateway as Unity AI Gateway, wrapping its existing governance model around agent access to LLMs and MCP servers. The A2A Protocol crossed 150 supporting organizations in April, and the Linux Foundation's Agentic AI Foundation now stewards both MCP and A2A.

What Does This Change for Builders?

I've spent the last year wiring agents to enterprise systems, and the integration tax was brutal: every platform wanted its own API wrapper and its own auth dance, and the glue code rotted faster than I could write it. That era is ending.

The practical shift:

  • Tool discovery is standardized. Your agent discovers tools the same way whether it's talking to Salesforce, Databricks, or your internal services. No custom code per platform.
  • Governance follows the protocol. Permissions, auditing, and policy enforcement wrap around MCP calls at the gateway level, so you stop building security per integration.
  • Composability is real. One agent run can chain tools across platforms - query CRM data, run a model, write the results back - through the same protocol layer.
MCP layerA coding-agents box at the top connects down into a highlighted MCP box that sits inside a dashed gateway container; three arrows lead from the MCP box down to Salesforce, Databricks and internal services. Coding agentsClaude Code, Cursor, CodexGateway (e.g. Unity AI Gateway):permissions, auditing, policySalesforce Headless 36060+ MCP toolsDatabricksYour internal servicesMCPone tool-discovery protocol97M monthly SDK downloads tool calls query CRM data,write results back run a model same discovery,no custom wrapper
One agent, one protocol: coding agents reach Salesforce, Databricks and internal services through MCP, with permissions, auditing and policy applied once at the gateway instead of per integration.

The second-order effect is the one I care about. Once every platform speaks MCP, your agent's capability surface grows every time someone else ships a server, whether or not you did anything. Integrations start arriving on their own.

Fragmentation, Then Convergence

Platform shifts tend to run through fragmentation first, then protocol convergence, and only after convergence does the higher-level capability take off. HTTP went through this arc for the web, REST for services, and MCP is at the same point for agents. We're past the convergence step now, and the debate about which protocol wins is over. The interesting work - the part I'm focused on - is what you build when the plumbing is no longer the problem, like agents that improve from their own execution traces.

Comments