MCP servers & connections
MCP (Model Context Protocol) lets agents and chat call external tools. Connections (when Automations are enabled) link external accounts such as Slack or Gmail via Pipedream Connect.
MCP servers (Admin)
Open Integrations → MCP servers.
| Task | What to do |
|---|---|
| Add remote HTTP MCP | Create a server with Streamable HTTP transport and endpoint URL |
| Install from catalog | Browse catalog (GitHub or custom images) → install |
| Import JSON | Paste Claude Desktop / Cursor-style mcpServers JSON (stdio/npx entries need Catalog → Custom) |
| Sync tools | Discover tool list after credentials are available |
| Assign to projects | Limit which projects may use the server |
Credential types include none, passthrough, OAuth, PAT, and env vars. End users connect OAuth/PAT under integrations when the server requires a personal credential.
Managed instances
When Kubernetes hosting is enabled, some catalog installs run as isolated worker pods (for example WhatsApp pairing). Use instance lifecycle actions (restart, suspend, pairing QR) from the MCP UI. Without hosted runtimes those APIs return unavailable.
Connections (Automations)
Open Integrations → Accounts when the automations feature is on.
- Start Connect for an app (Pipedream) for your project.
- Complete the vendor OAuth UI.
- Record the connection in Kroov — copy the webhook secret immediately (shown once).
- Use the connection’s webhook URL with Automations; rotate the secret if it leaks.
Tips
- Prefer Streamable HTTP over deprecated OpenAPI bridges.
- Never put long-lived vendor secrets in browser code — use the vault / Connect flow.
- Constrain tools with project assignment + per-user credentials + agent/automation grants.
- Connect your personal AI tools (Claude Desktop, Cursor, VS Code, …) to your own durable memory via the personal MCP server — see Personal MCP — connect your AI tools.