Skip to main content
Register an MCP server at the workspace level before granting it to an Agent Console agent. You can connect a custom endpoint or enable MCP for a supported native integration.
This workflow connects external MCP servers to Sidecar agents. To connect the Sidecar MCP server to an external AI client, see Use Sidecar MCP in an AI client.
You must be an Account Admin to manage MCP servers.

MCP setup guides

dbt Cloud MCP

Give agents dbt Cloud investigation and retry tools.

Snowflake MCP

Use Sidecar’s native Snowflake MCP setup.

BigQuery MCP

Use Google’s managed BigQuery MCP server with your integration identity.

Databricks MCP

Use the managed SQL MCP server from your Databricks integration.

Redshift MCP

Host or bridge a Redshift MCP server for Sidecar.

Requirements for a custom endpoint

Sidecar’s custom connector expects the exact MCP endpoint URL to:
  • Support the MCP Streamable HTTP transport
  • Be reachable from both Sidecar’s infrastructure and Anthropic’s Agent Console runtime
  • Accept a static token in the Authorization: Bearer <token> header
  • Support MCP initialization and tools/list
  • Respond to connection checks within 30 seconds
Use HTTPS for every production endpoint. A local stdio command, an SSE-only URL, or a server that requires an interactive OAuth flow cannot be registered directly.

Configure network access

For a direct MCP endpoint, allow both clients through its firewall, IP access list, and authentication policy: Check Anthropic’s current outbound addresses before configuring access. A local MCP Inspector test or a Sidecar connection test does not prove that Anthropic can reach the endpoint. Native Snowflake, BigQuery, and Databricks MCP connections go directly from Anthropic to the provider. Native dbt Cloud MCP goes through Sidecar’s proxy; Sidecar makes the upstream dbt Cloud requests. A hosted Redshift gateway needs access from both clients, while the gateway’s own identity and network govern its access to Redshift. Follow the provider-specific guide for network-policy and credential details.

Add a server

  1. In Sidecar, select Integrations.
  2. Open the MCP Servers tab.
  3. Click Add MCP.
Add an MCP server dialog with native connector and custom server options

Choose a native connector or add a custom MCP server.

Choose a setup method.
Select Custom server, then enter:
  • Display name — the label people see in Sidecar
  • Name — a stable machine-readable name
  • Server URL — the full Streamable HTTP MCP endpoint
  • Bearer token — the workspace credential used for Sidecar discovery and Agent Console runtime access
Click Add server.
Custom MCP server form with example display name, machine name, HTTPS URL, and bearer token

Use a stable name, the complete HTTPS endpoint, and a write-only bearer token.

Native BigQuery and Databricks MCP servers are managed from their warehouse integrations. You do not register them with a separate bearer token. Use the integration page even if BigQuery appears as coming soon in the Add MCP dialog. Redshift uses the hosted custom-endpoint workflow. Follow its guide for gateway and database access.

What happens when you save

For custom, Snowflake, and dbt Cloud setup, Sidecar validates the supplied token with authenticated MCP discovery before registration. It stores the token as a write-only workspace credential. The token is never displayed again. If registration fails, resolve the reported error and retry. Native BigQuery and Databricks use managed OAuth credentials from their linked integrations. Sidecar handles access-token renewal. Check the registered row’s status and tool catalog, then verify a tool call in a new Agent Console run.

Manage a server

For servers with a static bearer credential, use the row actions to:
  • Check now — check endpoint reachability without retrieving the stored credential
  • Refresh tools — re-enter the token and request the tool list again
  • Replace token — replace the write-only credential and repeat tool discovery
  • Edit — change a custom server’s display name or URL
  • Delete — remove the server from the workspace
For static-token connections, Check now performs an unauthenticated endpoint check. It does not validate the stored token, prove Anthropic’s access, or establish permission to query data. A Connected badge can retain an earlier authenticated result after a runtime authentication failure. Re-enter the token with Refresh tools to test discovery from Sidecar, then verify a tool call in Agent Console.
Native BigQuery and Databricks Check now actions use the linked integration identity for authenticated checks. Their credentials are managed on the warehouse integration’s Connection tab; manual MCP token replacement is unavailable. Even an authenticated discovery check does not prove access to every table or tool operation. Changing a custom server URL requires you to re-enter its bearer token. Native server URLs are managed by their linked integrations and are not editable on the MCP server row.
Deleting a server removes its agent grants, job selections, and tool-permission exceptions.

Give an agent access

Confirm the server is granted to the intended agent and selected for the job. Continue with Give agents MCP access, including its runtime verification steps.