Framework · AI & design · 3 min read
MCP
Model Context Protocol: an open standard for connecting AI agents to tools and data (design files, browsers, issue trackers), so you can decide which systems your team's agents may read or change.
The AI tool your designer works with. It decides which connector to call.
Typically: a coding or design agentA connector that gives the agent tools to read or act on one system.
Typically: Figma MCP, Playwright MCPThe system behind the server, with its own data and permissions.
Typically: design files, browser, trackerEach server = one more system the agent can reach · approve it first
Use it when
- Designers want an agent to read Figma frames, variables and component mappings instead of copying screenshots into a chat.
- You want an agent to open a prototype in a browser, take screenshots and check behaviour.
- Someone asks "can I connect our issue tracker or wiki to the agent?" and you need a clear answer.
How to use it
- Start from the task, not the server. Write down which system the agent needs and why: "read the Figma frame and its variables", "screenshot every state in a browser". Output: a short list of task → system pairs.
- Check the approved list. Ask IT security which MCP servers are allowed, and add the decisions to your AI usage guidelines. Extensions and MCP servers are a supply-chain risk, so install only from official sources. Output: approved servers, with who maintains each one.
- Limit what it can do. Prefer read-only access where you can. Know how each server signs in (personal token or single sign-on) and which data it can reach. Writing to a design file or tracker deserves the same care as any action with Human-in-the-loop approval. Output: a note per server: read or write, which data, which login.
- Use it in a simple loop. A typical design loop: read the Figma frame through the Figma MCP server, build with real design-system components, then open the result with a browser server (for example Playwright) and compare screenshots. Output: a prototype and a list of differences from the design.
- Share the setup. Write the working setup into the team playbook, or package it for the team, so everyone uses the same approved servers. Output: one playbook entry that others can follow.
Worked example
An internal operations platform with 6 product teams. A designer wants to check whether a built screen matches the design.
| Step | Server | What the agent does |
|---|---|---|
| Read | Figma MCP server (approved, read access) | Lists the components and variables used in the frame |
| Build | None, works in the prototype repository | Builds the screen with design-system components and anonymised mock data |
| Verify | Playwright MCP (approved) | Opens the prototype, screenshots desktop at two widths, lists differences from the frame |
| Report | Issue tracker server (requested) | Not on the approved list yet. The designer pastes the list manually and asks IT security to review the server |
The designer gets a "design vs built" list in minutes, and the unapproved connector stays out until it has been reviewed.
Watch out for
- Installing from anywhere. Unofficial extensions and servers can be malicious. Use official sources and your approved list.
- Write access by default. An agent that can write to the design file or the tracker can also make mistakes there. Start read-only and add write access on purpose.
- Personal tokens with broad access. A token tied to one person may reach far more data than the task needs. Check the scope with IT security.
- Assuming features and plans are stable. As of late 2026, what each server can do (and which plan it needs) changes often. Check the vendor docs before you promise a workflow.
Template
Copy this into your approved tools list, one block per server.
MCP server: Source (official link):
Used for (task):
Access: Read only / Read and write
Data it can reach:
Sign-in: Personal token / Single sign-on
Plan or seat needed:
Approved by: On:
Maintained by:
Allowed data classes:
Notes / limits: