n8n's analyst argues MCP moved integration work to vendors, but APIs, hooks and SDKs still carry much of the load
A post on the n8n blog revisits where integration work sits now that most vendors ship an MCP server.
What happened
The author of n8n's AI Agent Development Tools report argues that MCP shifted a real share of integration work from the API consumer to the vendor. In the old model the client handled auth flows, retries, data mapping and breaking changes for every integration. With MCP, the vendor publishes a server and has to keep it current against a spec that moves quickly. The post notes that recent revisions made the protocol stateless with per-request capability discovery, and that features get deprecated between versions.
The post makes several points that matter in practice. A valid tool definition doesn't mean an agent knows the workflow, so vendors now ship skills, markdown files that tell the agent which tools to call in what order and how to handle edge cases. MCP still sits on APIs: the specification has an MCP server act as an OAuth 2.1 resource server. Hooks are deterministic scripts the agent harness runs at fixed points, such as before a tool call, and the author assigns them to security and platform teams for policy checks and audit logs. SDKs like the Claude Agent SDK let developers define tools in the same codebase as the agent.
The conclusion is that the share of MCP compared with direct APIs, hooks and SDKs will become more balanced, because teams can't afford to spend tokens on every tool call.
My take
This matches how I split client builds. When a step is known in advance, like creating a contact, moving a deal stage or posting a Slack alert, I use a direct API call in a deterministic workflow step. It's cheaper, it's easier to retry safely, and it behaves the same way every run. MCP earns its place where an agent genuinely has to choose which action fits, such as triaging an inbound request.
Two of the post's points I'd add to any agent build now. Write a short skill file per integration describing the order of calls and the edge cases, because the tool schema alone won't tell the agent that a contact must exist before a deal is attached. And put policy in a hook or a gateway that runs before tool calls, since that check is deterministic and the model's judgment is not.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
