MCP in production: 3 integrations that paid for the setup

MCP (Model Context Protocol) graduated from Anthropic experiment in 2024-2025 to practical standard in 2026. Today OpenAI, Google, and dozens of vendors implement or route to MCP. It’s the “USB-C of agents” — protocol connecting LLM to tools without each pair needing custom integration.

Three integrations running in production at our clients.

What MCP is, quick refresh

MCP defines how model (client) and server (exposing tools, resources, prompts) communicate. You write one MCP server for an API; any MCP-speaking LLM can consume it.

Practical benefit: stop writing custom “tool wrappers” per model. An MCP-speaking server serves Claude, compatible ChatGPT, compatible Gemini, local agents.

Open spec: spec.modelcontextprotocol.io.

Integration 1 · MCP server for Bitable / Notion / Airtable

Scenario: company lives in Notion-like database. Team needs the agent querying and updating records.

Implementation:

• MCP server runs as local process or container.
• Exposes tools: list_records, get_record, update_record, create_record.
• Each tool validates schema before calling the API.

Where it pays:

• Agent becomes team’s “natural Notion.” “Who owns feature X?”, “what’s the status of project Y?” answered with source.
• Update via conversation: “mark task Z as done” is a command, not 3 clicks.

Gotcha: write governance. Don’t grant update permission without durable pause on critical tasks.

Integration 2 · MCP server for the ERP

Scenario: company with vertical ERP (Totvs, SAP B1, NetSuite). Want an agent answering operational questions without users entering the ERP.

Implementation:

• MCP server exposes read-only tools over ERP views (inventory balance, order status, monthly commission).
• Doesn’t expose write tools — writes still happen via official UI.
• Auth via service account with minimum needed permissions.

Where it pays:

• Salesperson asks “what’s product X’s margin?” in chat and gets it.
• Director asks “what’s AR balance this month?” without opening 3 screens.

Data-protection gotcha: ERP financial data is sensitive. Auth, log, and ACL critical.

Integration 3 · MCP server for CRM

Scenario: B2B company with Salesforce/HubSpot/Pipedrive. Salesperson wants briefing before call.

Implementation:

• MCP server exposes: get_account_summary, get_recent_interactions, get_open_opportunities, get_last_call_notes.
• Combines with web research tool (LinkedIn, news) also via MCP.
• Agent composes pre-call briefing of 1 page.

Where it pays:

• Salesperson enters call pre-warmed in 2 minutes instead of 20.
• Post-call: agent summarizes, writes note, updates stage (with human confirmation).

Gotcha: CRM data can be stale. Briefing must show last interaction date so salesperson calibrates confidence.

Why MCP beat traditional RPA

Classic RPA (UI automation) is fragile — breaks when UI changes. MCP is API-first — breaks only when the API changes, easier to detect.

Additional advantage: MCP is semantic. The agent knows what each tool does via schema + description. RPA only “clicks here.”

Why MCP beat custom direct integration

Before, each model + tool integration required custom code. MCP standardizes. You write the server once, the protocol handles the rest.

Historical analogy: HTTP beat proprietary protocols because any client speaks HTTP. MCP is becoming that for agents.

When NOT to use MCP

• Super-simple case (one API call per interaction). Direct function calling suffices.
• When latency is critical (MCP adds ~50-200ms overhead). For hard real-time, avoid.
• When the model vendor doesn’t support it. In 2026, becoming rare, but check.

Where to go deeper

For the protocol itself: spec.modelcontextprotocol.io and Anthropic examples. For the governance pattern around MCP in production (auth, logging, harness), see Harness Stack and multi-channel gateways applying MCP.