Integrate MCP servers into Claude Code and agent workflows

Put every MCP server in the project config so the whole team gets the same setup.

Personal and experimental servers then land in everyone's environment and in version control, and a broken experiment becomes a team-wide failure.

Project scope is for shared team tooling; user scope is for personal and experimental servers. Both are active at once.

Put the token in the config and add the file to .gitignore.

The secret still exists in plaintext on every machine, and one commit that ignores the ignore file publishes it.

Reference the variable and let it expand at load time; the value never enters the file.

Build a custom server for a standard system so it fits the team exactly.

You inherit maintenance of an integration someone else already maintains, and it drifts behind the upstream API.

Use a community server for standard integrations; reserve custom servers for genuinely team-specific workflows.

Register the MCP server and expect the agent to prefer it over the built-in tools.

With a thin description the model falls back to what it knows — searching with a built-in rather than calling the richer tool that would have answered directly.

Describe an MCP tool's capabilities and outputs in enough detail that it out-competes the generic built-in.

Model everything the server offers as tools — resources are an optimisation.

Without a catalogue the agent discovers what exists by calling tools speculatively, paying a round trip per guess.

Resources expose what is available; tools act. A catalogue removes the exploratory calls entirely.