On October 15 I watched Claude Code build two changes it had been told not to build yet. It had read the two review comments through the Skyelight MCP, and both had a hand-off set to plan: write up what you'd do, and change nothing. It built both in one turn. When I asked why, it said: “Nothing stopped me from seeing it. It was in the data, and I didn't act on it.”
The field was there. What the field meant wasn't. Our tool descriptions never said, so the agent read “plan” as a label. We rewrote the server's instructions that afternoon to say a hand-off outranks a go-ahead in chat.
That's the part most write-ups of the Figma MCP, the Playwright MCP and the Chrome DevTools MCP skip. Each server gives your agent one thing it couldn't reach before: the design, a browser, the backlog. What the agent does with it depends on which servers you connected, and on how well each one explains itself.
What is an MCP server?
It's a program that offers your agent a set of tools. Anthropic introduced the Model Context Protocol in November 2024, and in December 2025 handed it to the Agentic AI Foundation, part of the Linux Foundation. When Claude Code, Cursor or Codex connects to a server, it asks for the tool list: a name, a description and the inputs for each. The agent reads those descriptions and decides on its own when to call what. Some servers run on your machine with npx. Others are hosted, and you sign in once.
Figma MCP: the design
Figma's server hands the agent the frame you're building from. get_design_context returns the layout and styles of a selection, get_variable_defs the tokens behind it, and get_screenshot a picture to check against. With Code Connect set up, get_code_connect_map tells the agent which of your components a Figma component already is, so it reuses Button instead of writing a new one.
The hosted server at mcp.figma.com/mcp is the one Figma recommends, and only it can write back: turning a live page into design layers, or making a new file. The desktop server needs the Figma app open and a Dev or Full seat on a paid plan. Watch the limits. A View or Collab seat gets 20 calls a month, which one afternoon of real work will use up.
Use it when there's a design file and the build should match it. If the app was prompted into existence and there's no frame, skip it.
Playwright MCP: a browser to test in
Microsoft's server gives your agent a browser it drives itself. It reads the page as an accessibility tree, the same structure a screen reader uses, so it doesn't need a vision model to find the sign-up button. browser_navigate, browser_click, browser_fill_form and browser_console_messages cover most of a test run. It runs locally with npx @playwright/mcp@latest, and with the Playwright extension it can use the tabs you're signed in to.
Two notes from its own README. It says plainly that it's “not a security boundary”, so treat anything on a page it visits as text your agent might follow. And it suggests coding agents may do better with the Playwright CLI and skills, which use fewer tokens.
Use it when the agent has made a change and should walk the flow before telling you it's done.
Chrome DevTools MCP: why it's broken or slow
Google's server is DevTools for your agent. Where Playwright asks “does the flow work?”, this one asks why it doesn't. list_network_requests and list_console_messages show what failed, performance_start_trace records a trace the agent can read back as findings, lighthouse_audit returns the failing audits, and get_css_styles settles which rule won.
It can attach to the Chrome you already have open. Start it with --autoConnect on Chrome 144 or later, allow remote debugging, and the agent sees your signed-in session instead of a fresh browser that can't get past your login. It has 61 tools, which is a lot of descriptions to load into every session. The --slim flag cuts it down.
Use it when the question is a slow page, a failing request or a style that won't apply.
Linear MCP: the backlog
Linear hosts its server at mcp.linear.app/mcp. You sign in with OAuth, and the agent can find, create and update issues, projects and comments. There's a read-only address at /mcp/readonly for agents you only want reading. On November 5 Linear started sending clients default instructions on how to use its tools, which is the same lesson we learned in October, from the other side.
Use it when the agent should start from the issue, or file one for something it found and won't fix now.
Skyelight MCP: what the reviewer thought
None of the four servers above knows what your client said about the pricing page. That lives in a review: a comment pinned to the element, a reply or two, sometimes a decision. Skyelight's server gives your agent those threads. list_items shows what's open, get_item returns one thread with the element it's on and, with Skyelight Build, the file and line that render it. post_update puts the agent's answer back on the thread, where the reviewer will see it.
The hand-off is the part we got wrong first. Set a thread to investigate or plan and the agent replies on the thread and changes no code, even if you tell it to in chat. Set it to fix and it builds. It's hosted at app.skyelight.ai/mcp with OAuth, and Free includes all 18 tools, with 25 agent threads a month.