Integrations overview

Plugins connect Claude to external tools through MCP servers bundled in a plugin's .mcp.json, and through one command-line tool. An MCP server adds nothing to the skill listing, and Claude Code defers its tool schemas until a tool is used, so a session that never touches a server pays only for its tool names.

What ships where

Integration	Bundled by	Transport	Sign-in or prerequisite
Atlassian	tsh-core	http	OAuth via /mcp
Figma	tsh-core	http	OAuth via /mcp
Context7	tsh-core	http	None
AWS Documentation	tsh-stack-aws	stdio (uvx)	uv installed
GCP: gcp-gcloud, gcp-observability, gcp-storage	tsh-stack-gcp	stdio (npx)	See the page
Playwright CLI	Not bundled; a CLI you install	Command line, not MCP	npm install -g @playwright/cli@latest

tsh-stack-azure, tsh-platform-engineering and tsh-product-testing bundle no server of their own.

One home per server

A server lives in exactly one plugin. Plugin-provided servers are deduplicated by endpoint, but the surviving definition decides the tool namespace, and the winner tracks plugin load order, which nobody controls. That is why tsh-product-management, tsh-product-engineering and the other plugins that read Jira or Figma depend on tsh-core for the servers instead of shipping a second copy.

Cloud servers ship with the cloud's own plugin: a GCP repository never starts an AWS server, and no cloud server appears in a repository with no cloud plugin installed.

Signing in

Run /mcp once per machine and sign in to atlassian and figma. Step by step: Connecting tools.

Do not pin tool names

Tool names depend on which server entry wins. A teammate's own user-scope figma entry takes precedence over the bundled one, and the namespace becomes mcp__figma__* rather than mcp__plugin_tsh-core_figma__*. The plugins therefore never name a server's tools in an agent's tools:, a skill's allowed-tools: or a hook matcher, and discover what the session exposes at runtime instead.