Building a creative team with Claude Code and MCP

A creative team that feeds paid media needs to do a lot more than make ads. It has to research customers, write briefs, track what’s been tested, organise assets, pull performance data and turn results into the next round of ideas. In most small teams, that operational work quietly eats the time that should go into creative thinking.

Over the past few months we’ve rebuilt much of that work around Claude Code, Anthropic’s agentic tool that runs in the terminal, and the Model Context Protocol, or MCP. The result isn’t a replacement for creative people. It’s a set of assistants that handle the groundwork, so the people can focus on judgement and craft.

The two pieces

Claude Code is an AI agent that works inside a project folder. It can read and write files, run scripts and carry out multi-step tasks with you reviewing along the way. It’s built for software work, but anything that lives in files, such as briefs, copy, spreadsheets, scripts and reports, can benefit.

MCP is an open standard, introduced by Anthropic in late 2024, for connecting AI tools to other systems. An MCP server exposes a tool or data source, such as a document store, a design tool or a database, in a way the AI can use. Instead of copying information in and out of a chat window, the agent can look things up and take actions directly, within the permissions you set.

Together, they let us give an AI assistant real context about a brand and real access to the tools the team already uses.

One workspace per brand

Each brand gets a project folder that acts as the team’s shared memory:

• Brand guidelines and voice, written for a reader who needs to apply them, not admire them.
• An approved claims list, so nothing makes it into an ad that the brand can’t substantiate.
• Customer research: reviews, survey answers, support themes and the language customers use.
• The test log: every concept tested, the hypothesis, the result and what we learned.
• A CLAUDE.md file, which Claude Code reads automatically. It summarises the brand, where everything lives and how we like work done.

Because the context lives in files, every task starts informed. The assistant doesn’t need the brand explained again each time.

Where it helps most

Research. We give it a large export of reviews and ask for themes, objections, outcomes customers describe, and memorable phrases. It does in minutes what used to take an afternoon, and it cites the reviews it’s drawing from so we can check.

Briefing. Given the research, the test log and a concept direction, it drafts briefs in our standard format, including hook options, proof points and the hypothesis. A strategist edits and approves every one.

Copy variations. For winning concepts, it produces headline, primary text and hook variations within the claims list and character limits, written straight into a sheet ready for review.

Reporting. Connected to performance data through MCP, it summarises weekly results by concept, flags fatigue, and updates the test log with what happened. It turns dashboards into sentences the team can act on.

Asset admin. Naming conventions, folder structures and spec checks against each placement’s requirements. Small tasks, but they add up.

Guardrails we keep

Giving an agent access to real tools calls for care.

• Read access first. Most connections are read-only. Anything that changes live systems, especially ad accounts, needs human approval.
• Humans approve creative. Every brief, script and line of copy is reviewed by a person before it’s used.
• No invented proof. The assistant works from the claims list and real reviews. It doesn’t make up testimonials or results, and we check that it hasn’t.
• Version control. Changes to briefs and documents are tracked, so we can see what changed and why.

What it changes

The biggest shift isn’t speed, though things are faster. It’s that the unglamorous work now gets done consistently. Every test gets logged. Every week gets summarised. Every brief draws on the full history of what the brand has learned. A small team with good judgement, supported by well-set-up AI assistants, can now run a creative operation that would have needed several more people not long ago.