Plan the deployment
The AWS Labs setup starts the server as a localstdio process. Sidecar cannot
run that command. To use it with Agent Console, deploy it behind a service or
bridge that provides:
- An HTTPS MCP Streamable HTTP endpoint
- A stable bearer token for Sidecar
- Access to AWS through an IAM role or another AWS credential-provider-chain method
- Network access to the Redshift Data API and any required private resources
https://redshift-mcp.example.com/mcp. It
must meet Sidecar’s custom endpoint
requirements.
Configure gateway network access
Sidecar’s registration and health checks reach your hosted MCP gateway from44.236.246.98/32. Agent Console calls the same gateway from Anthropic’s
infrastructure. Permit both sources at the gateway’s firewall and
authentication layer. Use the current outbound
addresses.
The gateway connects onward to AWS and Redshift with its own workload identity
and network. Allowing Anthropic at the gateway does not grant it direct access
to your Redshift database. A successful local or Sidecar discovery check does
not prove that the runtime can reach the gateway.
Configure AWS access
Attach a dedicated IAM role to the hosted MCP workload. The AWS Labs server documents these permissions as its starting point:AWS_REGION=us-east-1. Prefer an attached IAM role over static
AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY values.
Configure database access
IAM permission to call the Redshift Data API does not grant access to data. Create a least-privilege database identity withUSAGE on allowed schemas and
SELECT on allowed tables and views.
The Redshift warehouse integration guide
includes SQL for a read-only sidecar_user, schema grants, future-table grants,
and a grant-statement generator. Reuse it only when the MCP server should have
the same access as Sidecar’s metadata integration.
Avoid granting CREATEUSER solely for the AWS Labs review_cluster tool. The
AWS Labs guide notes that this tool needs elevated system-view access. Omit or
filter it at your hosted MCP layer when agents only need metadata discovery and
read-only queries. If you cannot filter it, leave the elevated privileges
ungranted so the tool cannot perform the review.
Validate the server before hosting
Use the official local configuration to confirm AWS and Redshift permissions before adding the HTTP bridge:Register the hosted endpoint
- In Sidecar, select Integrations → MCP Servers.
- Click Add MCP, then select Custom server.
- Enter values like these:
- Click Add server.
- Confirm that the server status is Connected and review the discovered tools.
Verify Agent Console access
After granting the hosted server and enabling it for a job, start a new manual run. Ask the agent to use the Redshift MCP query tool forSELECT 1 and a
SELECT from one permitted table with LIMIT 1. Supply the exact cluster or
workgroup, database, schema, and table.
Expand the MCP calls in the run transcript. Confirm successful query results
from the intended database, and complete any required approval. An empty table
can return zero rows successfully. See runtime
verification.
Troubleshooting
Sidecar reports the server as unreachable
Sidecar reports the server as unreachable
Confirm that you registered the public or privately routable HTTPS
/mcp
endpoint. A local uvx command or localhost URL is not reachable from
Sidecar. Check the gateway’s TLS certificate, firewall, and request timeout.Connected in Sidecar, but the agent has no Redshift tools
Connected in Sidecar, but the agent has no Redshift tools
Confirm the agent grant and job selection, then start a new run. Check that
the gateway permits Anthropic’s outbound requests and accepts the stored
bearer credential. A Connected badge can retain an earlier Sidecar result.
Send support the run URL, timestamp and time zone, and sanitized gateway
error if initialization still fails.
The server connects but cannot find clusters
The server connects but cannot find clusters
Check the workload’s AWS region and IAM permissions for provisioned Redshift
and Redshift Serverless discovery. Confirm that the attached IAM role is the
identity the server is actually using.
Queries fail after clusters are discovered
Queries fail after clusters are discovered
Check the Data API credential actions and the Redshift database user’s
USAGE and SELECT grants. Cluster discovery and database query access use
separate permissions.