Your AI agent is only as useful as the data it can access. Most business data lives in a CRM, a database, a spreadsheet, or an internal tool. MCP is how you connect your agents to that data without rebuilding the integration for every new AI product you adopt.
The integration tax
Every time you want a new AI tool to access your business data, someone has to write an integration. OAuth flows, API wrappers, data transformation, error handling. Multiply that by every data source and every AI product, and the integration tax becomes the reason most enterprise AI projects never ship.
MCP doesn't eliminate integrations, but it eliminates the duplication. You write a Salesforce MCP server once. Every MCP-compatible agent — your Claude setup, your custom assistant, your future tool — can use it. The integration becomes infrastructure instead of overhead.
What MCP gives agents access to
A well-designed MCP setup gives your agents the ability to:
Read — query your CRM for customer data, pull from databases, read documents, check inventory, retrieve past conversations.
Write — create records, update statuses, send messages (with appropriate guardrails), log decisions, generate reports.
Act — trigger workflows, schedule events, initiate processes in external systems.
The combination of read, write, and act is what separates useful business agents from expensive chatbots.
The MCP servers worth knowing about
The ecosystem has grown fast. Servers that exist today and are production-ready:
Data and databases: PostgreSQL, SQLite, MySQL, BigQuery, Snowflake. Connect your agent to any database and let it write queries.
Productivity: Google Drive, Google Sheets, Notion, Obsidian. Document reading, writing, and organization.
Communication: Slack, Gmail, Linear, GitHub. Agents that can read issues, send messages, and create tasks.
CRM and sales: HubSpot, Salesforce (community-built). Customer data lookup and enrichment.
Code and devtools: GitHub, GitLab, Jira. Engineering workflow automation.
Web: Browser automation, search, scraping. Agents that can navigate the web.
If your system isn't on this list, writing an MCP server for it is a half-day project for a competent developer.
How to think about which servers to build
Start by mapping the 20% of data sources that account for 80% of the questions your team asks. That's where you start.
For most businesses, the first three MCP servers are:
- Your main database (or data warehouse)
- Your CRM
- Your communication tool (Slack or email)
With these three, an agent can answer most operational questions: what's happening with this deal, what did we ship last week, what's the status of this customer.
Security is the real conversation
MCP doesn't automatically make your data safe to expose to a model. Before you connect anything:
Scope access tightly. The agent should only see what it needs for the task. A customer support agent doesn't need finance data. Build servers with fine-grained permission models.
Audit what's being sent. Log every tool call. Know what data your agent is reading and writing. You need this for compliance and for debugging.
Use read-only servers where possible. An agent that can only read data can't corrupt anything. Add write access incrementally, not all at once.
Keep humans in the loop for consequential actions. Deleting records, sending external communications, financial transactions — require human approval before execution. Build approval flows, not just permission systems.
Test with production-like data in a staging environment. Agents interact with data differently than humans. You will find edge cases. Find them in staging.
The compounding value
Every MCP server you build today is reusable infrastructure. When the next AI product arrives — and they're arriving fast — you don't rebuild integrations. You connect it to your existing servers.
This is the strategic reason to invest in MCP now. The businesses building this infrastructure today will have a meaningful operational advantage in 18 months. The integration tax they've already paid is amortized across every new AI tool they adopt.
The businesses that wait will pay it again and again.
Published under Field Notes. Not for sale. Share freely.