Release / / with 2026.10.20261002180809, 2026.10.20261012173232
Open source MCP servers for AWS 2026.10.20261015170211
Modify Redshift data with optional transactions that you can commit or roll back.
Before you upgrade to 2026.10.20261015170211
12 to fixInstall `boto3` and `botocore` >=1.43.26 when managing dependencies manually.
Set `DOCUMENTDB_CONNECTION_STRING` or `--connection-string` at startup; remove `connect`/`disconnect` calls, all `connection_id` arguments, and `--connection-timeout`.
Rename tool calls from `enterprise_support` to `enterprise-support`; read `estimated_total_deduped_savings` instead of `estimated_total_savings` in Cost Optimization Hub summaries.
Large `sp-explorer` offering, inventory and rate results are queried through `session-sql` instead of returned inline; small results remain inline.
Idle recommendations default to savings-after-discount descending. Compute Optimizer and Cost Optimization Hub normalize filter spelling and report structured validation errors.
Use `InstanceFinding` or `StorageFinding` instead of ambiguous RDS `Finding`; send at most one `account_ids` entry and cap `max_results` at 1000 (100 for idle recommendations).
Failures expose top-level `error_type`, `operation`, `service` and `message`, plus `retry_strategy` and optional `hint`; Cost Optimization Hub AWS errors no longer rely on nested `data.error_code`.
For tag values in `COST_AND_USAGE`, replace `getDimensionValues` with `dimension=TAG` by `getTagsOrValues`, optionally supplying `tag_key`.
Scope IAM privilege-escalation actions to specific resource ARNs; broad grants using `NotAction`/`NotResource` and principal-wide escalation grants are now rejected.
Install `boto3` and `botocore` >=1.43.90 when managing dependencies manually.
Use MCP SDK >=2.1.0 and <3.0.0 when managing dependencies manually.
Remove the `updatesNextToken` parameter and stop reading `planUpdates`/`updatesNextToken` output; use plan steps and `stepsNextToken` for manual pagination.
Pass `cluster_type` (`provisioned` or `serverless`) to every tool taking `cluster_identifier`. Add `redshift-data:BatchExecuteStatement`; without it only the read-only compatibility path works.
Replace `FASTMCP_LOG_LEVEL` with `LOG_LEVEL`; use MCP SDK >=2.1.0 and <3.0.0 for manually managed installs.
Results above `MAX_RESULT_ROWS` (default 1000) are refused rather than truncated. `AWS_PROFILE` takes precedence over environment credentials; logs default to stderr at `WARNING`.
What changed in 2026.10.20261015170211
Redshift adds opt-in `ACCESS_MODE=read-write` and named transactions, while existing calls now require `cluster_type`. Oracle and DocumentDB pin connections at startup; billing renames `enterprise_support` and a savings-summary field, and adds recommendation filters and SQL offload.
Also: Oracle Database@AWS and RDS tenants; HealthOmics tags and session policies; billing queries and error handling; connection and tool-call upgrade requirements.
Available in the awslabs MCP packages at these releases. Requires an MCP client plus each server's runtime, credentials and permissions; Redshift writes remain opt-in.
New things you can do
Connect to Autonomous and Exadata Dedicated databases.
Opt in with `ACCESS_MODE=read-write`; execute writes and DDL, optionally grouping statements in named transactions.
Fixed and improved
Connect to RDS Oracle tenant databases using startup `--tenant_database_name`, with managed-secret lookup when no custom secret is supplied.
Read Cost Optimization Hub enrollment and preferences, and Billing Conductor transfer preferences; request resource-specific or Graviton recommendations with `resource_arns` and `cpu_vendor_architectures`.
Group Cost Optimization Hub recommendations with `group_by='TagKey:<key>'` and filter by tags; inspect expanded Compute Optimizer risk, utilization, savings and configuration fields.
Retrieve RDS recommendation details, paginated tag values and correct UTC refresh timestamps; empty pagination tokens now mark the last page, and concurrent SQL offloads retain table registration.
Pass native lists/objects for Cost Explorer `metrics`, `group_by` and `filter`, as well as JSON strings; use trailing `Z` in Compute Optimizer Automation datetimes.
Apply tags to `StartAHORun` and inline `session_policy` to runs or batches; policies accept a JSON string/object up to 2048 characters and scope down the run role.
Continue genomic file searches with advancing tokens and reject unparseable tokens; lint WDL and CWL input written as UTF-8.
Read the evidence
Read failures now retain their reason at the MCP client; recoverable 4xx and empty redirected pages also report a documentation-search next step.
Read re-planned jobs without a hard-coded plan version; `get_resource(plan)` retrieves all step pages, and jobs without plans return `NOT_FOUND`.
Run concurrent queries and use named read-only transactions for a consistent snapshot across calls; read complete result sets rather than only the first page.
Seeded by 