$ man context-wiki/mcp-servers

Model Context Protocol is the bridge between your AI agent and your production tools. Without MCP, Claude is blind. It can read files in your repo and run terminal commands, but it cannot see Slack messages, pull campaign data from HeyReach, or push drafts to Substack. It is stuck inside the code editor. With MCP, Claude has hands. Each MCP server connects Claude to an external tool and exposes a set of actions the agent can take. Read a Slack channel. Export leads from HeyReach. Create a draft on Substack. Deploy to Vercel. MCP turns Claude from a coding assistant into an operating system.

PATTERN

Slack MCP: read channels, send messages, search message history. When I run /slacksync, Claude reads the partner Slack channel and extracts action items, decisions, and deliverables. HeyReach MCP: export leads, track connections, pull campaign stats. When I run /heyreach-export, Claude pulls accepted and unaccepted connections and generates CSVs. Substack MCP: push draft posts. When I run /finalsubstack, Claude formats the content and creates the draft directly in Substack. Browserbase MCP: browser automation for LinkedIn recon. Claude visits LinkedIn profiles, extracts recent posts, and generates personalization hooks. Vercel MCP: deploy, check build logs, monitor deployments. The /deploy skill uses this to confirm sites are live after pushing.

CODE

MCP servers are configured in your Cursor settings or Claude Code config. Each server needs: a name (how the agent references it), a command to start the server (usually npx or node), and any authentication tokens (API keys, OAuth tokens). The server starts when you launch a session. It runs in the background. The agent sees the server's tools as available actions in its tool list. You do not call MCP directly. You use skills or prompts that trigger MCP actions. The agent handles the actual API calls. Installation is usually one config block plus an API key. The /addmcp skill automates this for common servers. You type /addmcp slack and it researches the server, adds the config, and verifies the connection.

PRO TIP

Before MCP, my workflow was: open Claude, write content, copy it, switch to Substack, paste it, format it, publish. With MCP: /finalsubstack and the draft appears in Substack. Before MCP: export HeyReach data manually, download CSV, open in Sheets, filter. With MCP: /heyreach-export accepted and the CSV is generated in my repo. Every manual step between AI output and production action is friction. MCP eliminates that friction. The agent goes from generating output to taking action in one step. That is the difference between AI as a tool and AI as an operator.

ANTI-PATTERN

MCP servers fail for predictable reasons. Authentication expired: API tokens have TTLs. If the server suddenly stops working, check the token first. Rate limits: hitting the external API too fast. Add delays between calls or reduce batch sizes. Server not starting: the command in your config is wrong, or a dependency is missing. Check the config path and run the command manually to see the error. The biggest anti-pattern is adding MCP servers you do not use. Each server adds startup time and context. Only install what you actively use in your workflows. Five focused MCP servers beat fifteen unused ones.