The MCP servers we actually use
A year and a half into the Model Context Protocol, here’s the set of connections that earn their place in our day-to-day work, and the security rules that come with them.
When we first connected AI agents to our tools through the Model Context Protocol, we connected everything we could. That was a mistake. More connections meant more noise, slower responses and more permissions to worry about. Over time we’ve trimmed down to a short list of MCP servers that genuinely change how the team works.
For anyone new to it: MCP is an open standard for connecting AI applications, like Claude, Claude Code or other agents, to outside tools and data. Each MCP server exposes a particular system, such as a document store, a design tool or an analytics platform, so the AI can look things up and take actions within the permissions it’s given. Many software companies now publish official servers, and the community has built many more.
The short list
Documents and knowledge. Our brand guidelines, briefs, research and test logs live in shared docs. Connecting them means every task starts with the right context, without anyone pasting it in. This is the single most useful connection we have.
Store data. A connection to the eCommerce platform lets the agent answer practical questions: which products sell together, how a price change affected conversion, which items are running low. It’s read-only for us.
Analytics. Site analytics through MCP turns “can someone pull the numbers on…” into a sentence the agent can answer in a minute, with the query visible so we can check it.
Email and SMS platform. Useful for reviewing flow performance, drafting new campaign copy within brand rules, and summarising what’s working across segments. Sending stays manual.
Ad performance data. Results by campaign, ad and concept feed our weekly reporting and the test log. There are several options here, some official and many community-built, and they vary a lot in quality. We use a narrow, read-only connection.
Design tools. Pulling specs, exporting assets and checking that creative matches the brand system. Handy for consistency, especially across many ad sizes.
Team chat. For delivering summaries where the team already works, so reports get read instead of sitting in a folder.
What we removed
We disconnected anything that wasn’t used at least weekly, anything that duplicated another connection, and any community server we couldn’t review properly. Each connection adds tools the agent has to choose between, and more choice doesn’t always mean better results.
Security rules
Connecting AI to real systems needs real care. Our rules:
- Least privilege. Every connection gets the narrowest permissions that do the job. Read-only is the default.
- No unsupervised writes to anything live. Ad accounts, the store, email sends: changes require a person to approve them.
- Vet before installing. We prefer official servers from the software vendor. Community servers get reviewed, and kept only if they’re maintained.
- Be wary of untrusted content. Text an agent reads from the outside world, such as web pages, reviews or emails, can contain instructions designed to manipulate it. Agents that read untrusted content shouldn’t also hold powerful permissions.
- Separate credentials. Dedicated, revocable keys for each connection, never shared personal logins.
- Log everything. We keep a record of what the agent accessed and changed.
What it adds up to
None of these connections is impressive on its own. Together they remove the constant friction of moving information between systems: copying numbers into reports, pasting guidelines into prompts, chasing the latest version of a brief. The team spends less time fetching and more time deciding. That’s the real value of MCP for a creative and growth team. It isn’t automation for its own sake. It’s context, available when it’s needed.