MCP Config

MCP Config is the JSON file that defines every external tool connection available to your AI host. Released alongside the Model Context Protocol in late 2024 by Anthropic it quickly became the standard way to declare which servers Claude Desktop, Cursor, Replit, and other hosts should launch at startup. The file lives at a specific path depending on your operating system. On Mac it sits in ~/Library/Application Support/Claude/claude_desktop_config.json. Windows uses a comparable path under AppData. Linux hides it in ~/.config/Claude. Cursor stores its version in a dedicated settings location but accepts the identical structure. The root object contains a single mcpServers key whose value is another object. Inside that you list each server by name with its launch command typically npx, an args array that usually starts with -y followed by the package name and any configuration flags, and an env object for secrets like personal access tokens. When the host starts it parses this file, launches each server as a background process, and registers the available tools with the model. The result is an AI that knows your actual Figma component names, your real GitHub issue list, and the precise content of your local design system documentation instead of whatever you bothered to paste in the chat that day. This single file eliminated the worst parts of AI design work by 2026.

MCP Config is not the MCP servers. Servers are the executable code that actually perform actions like reading a Figma file or creating a GitHub branch. The config merely tells the host which servers to run and with what credentials. It is not a graphical interface with switches and buttons. Cursor offers a partial UI but the source of truth remains the JSON. It is not self documenting. A year from now you will open the file and wonder why you added that weird env variable unless you comment it liberally. It is not forgiving of syntax errors. One missing comma or an extra quote and the entire mcpServers object fails to load with zero helpful error messages in the UI. It is not secure storage for tokens. The file sits in plain text on your hard drive. It is not a replacement for learning what each server can actually do. Pointing a server at the wrong directory or giving it insufficient permissions creates connections that appear in the tool list but return empty or erroring responses when called. It is not permanent infrastructure. You will edit this file every time your tool set changes or when a server updates its command line interface.

The concrete payoff shows up the moment you stop being the translation layer. Take product designer Priya at Notion in 2026. Her team ships updates to the shared design system every two weeks. Previously she would export JSON from Figma, copy Notion page content, take screenshots of the production site, and paste everything into Claude for review. The prep work alone took forty five minutes per session and she still had to answer follow up questions when the AI hallucinated details. After building her MCP Config that changed completely.

Her config file contained six server definitions. The Figma entry used the command npx with args pointing to the official figma-developer-mcp package and her read only API key passed as an argument. The filesystem server targeted her local design system folder containing token files in tokens.json and exported SVGs so the AI could compare intended versus actual asset usage. GitHub MCP received a token with contents read write scope limited to the design-system monorepo. The Notion server used a secret integration token with access to the workspace containing all project briefs. Linear MCP connected because her team tracked all system tickets there. The Playwright server allowed live DOM inspection of their public component gallery.

One Thursday afternoon Priya pasted a link to a new frame in the design system file and typed a single prompt. Audit this new icon set against our current guidelines and create tickets for every violation with proposed fixes. The AI immediately called the Figma MCP server to extract every icon variant, their layer names, color values, and corner radii. It cross referenced those values against the master tokens file on her filesystem. It pulled the current icon component code from GitHub to see what actually shipped. Three violations appeared. The stroke widths were inconsistent on two icons. One used a hardcoded color instead of the semantic token. The AI then used Linear MCP to create three separate tickets each populated with the exact frame link from Figma, a generated screenshot from the browser server showing the live icons, and a clear description of the fix including the exact token to swap in. It updated the original Notion page with a summary of findings and linked the new tickets. Priya watched the whole process unfold in real time. The session that previously took two hours of busywork now took twelve minutes and produced better results because the AI operated from source truth instead of her imperfect summary.

Freelance designer Maria maintains three separate MCP Config files for her clients. Client A is a fintech startup that lives in Figma and GitHub so those two servers get first position and generous permissions. Client B runs everything in Notion and Linear so those servers dominate her second config. The browser automation server remains constant across all three because competitive teardown sessions are part of every kickoff. Switching configs takes one copy paste into the right folder and a full restart. The time savings let her take on two additional clients without increasing her workload or context switching overhead.

Install your MCP Config the first time you plan to use AI as a daily collaborator rather than an occasional sounding board. Update it every quarter as new servers appear for tools like Slack, Google Drive, or your company internal wiki. Use it without hesitation on personal machines where you control the security boundary. The config shines brightest during design system audits, handoff preparation, competitive analysis, and ticket grooming sessions where accuracy beats vibes. Avoid MCP Config if your company security policy prohibits any local tool servers that connect to cloud models. Skip it for pure creative exploration where you want the AI to generate new ideas rather than analyze existing work. Do not use the same config across work and personal projects if that risks leaking client data. Never paste real production tokens into a config file that lives in a synced Dropbox folder or public repository. Test every addition by asking the AI a simple targeted question that requires that specific server before declaring victory. One broken server can prevent the entire MCP panel from appearing and send you down a JSON rabbit hole.

MCP Config is the unglamorous but essential foundation that turns your scattered design tools into a single coherent context the AI can finally act inside instead of endlessly guessing at.