Connectors & Pre-Built Integrations
Connectors are pre-built MCP servers for popular services: GitHub, Slack, Linear, Notion, Google Drive, Jira, Sentry, Stripe. Most are official (vendor-maintained) or from Anthropic's mcp-registry. Install via npm/pip or as plugins.
Connectors are app-store apps for agents — pre-built (GitHub, Slack, Linear), one-click install, vendor-maintained. Easier than rolling your own, but check permissions before granting.
Discovery: mcp-registry catalogs available connectors. Installation: usually npm install -g <package> or via plugin manager, then add to settings.json. Auth varies: API keys in env vars, OAuth flows (the connector handles the flow), service-specific tokens. Common patterns: read-only connectors (search, fetch) are safer; write-enabled connectors (post message, create issue) need more careful permissioning. Some connectors are server-side (hosted by the vendor), some local (run in your shell).
Choosing a connector: prefer official (vendor or Anthropic) over community for production work — auth flows, rate limiting, and edge cases are usually more polished. Read scopes only when possible — a GitHub connector with read-only scope can search/read but not modify. Connector stability varies — some break with API changes; lock to versions. Multi-tenancy: most connectors assume single-user; be careful with shared environments. Cost surprise: connectors that hit metered APIs (OpenAI, paid SaaS) can rack up bills if the agent loops on them — set rate limits.
Connectors are pre-built MCP servers for popular services — GitHub, Slack, Linear, Notion, etc. I install them like npm packages and configure auth in settings.json. The choice is between official and community connectors; for production work I stick to official because auth, rate limiting, and error handling are more polished. Read-only scopes when possible — a GitHub connector with read scope can answer questions about my repos without risk of accidental writes. Watch for cost surprises if the connector hits metered APIs.
Connectors with write access running unsupervised. An agent in a loop with a 'create Linear issue' tool can flood the issue tracker. Use minimal scopes and consider PreToolUse hooks for write actions.