Connect an AWS account
betaGive agents AWS API access controlled by your IAM role for conversations, scheduled checks, and event-triggered tasks.
A workspace admin can connect an AWS account using a customer-owned IAM role. Your AWS IAM policy determines the available operations. Authorized chats can use permitted writes and sensitive data reads. Scheduled wakeups require a separate explicit grant and are limited to reviewed infrastructure and cost reads. AWS access must be enabled for the workspace.
Connect from chat
Ask your agent to investigate AWS and name the account and Region. If the account is not connected, the agent can show a Connect AWS account card in web chat. A workspace admin can open role setup over the conversation. Check the account, Region and workspace before continuing. AWS access is separate from the connector catalog; you do not need to install a CLI or upload access keys.
The card resumes an existing unfinished connection. If several connections could apply, select the one to use. An already connected account offers Continue investigation. After successful verification, Verify and continue resumes the original investigation with that connection and read-only access. If the agent is busy or continuation fails, use Continue investigation when it becomes available; the connection is preserved. Revoked or unavailable connections require review in Cloud accounts.
If you are not a workspace admin, share the Cloud accounts link with an administrator. Account setup does not authorize a scheduled check. Scheduled checks still need their separate grant. An event-triggered task can also propose setup in the agent conversation. Other chat surfaces direct users to Cloud accounts rather than opening setup there.
Connect your account
- Open Settings → Infrastructure access → Cloud accounts, choose the workspace, and select Connect account. Select Amazon Web Services (AWS) and enter a connection name, 12-digit AWS account ID, and default AWS Region. Connections belong to the selected workspace. Old AWS settings links still open this page with the original workspace selected.
- Select Continue to role setup. Choose the role template permissions and All regions or Selected regions. Acknowledge application data reads or operational access if selected. Then choose CloudFormation, Terraform, AWS CLI, or AWS Console. CloudFormation provides a template and stack parameters; Terraform provides a configuration file; AWS CLI provides policy files and commands; AWS Console provides manual steps and policy JSON. Apply the role in the specified AWS account through your reviewed infrastructure workflow. Each method uses the connection’s d5s principal, external ID and selected role template permissions.
- Select Continue to verification, paste the resulting role ARN, and select Verify connection. Verification checks the account identity. Verification fails unless the trust policy requires the exact external ID shown. Use the supplied exact trust policy. The connection becomes Connected when those checks pass; AWS checks service permissions on each operation. If verification fails, check the account, ARN and trust policy, then retry. You can close setup and return through Details → Resume setup without creating another connection.
- In a web chat, an authorized Slack or Teams conversation, or trusted managed email, ask the agent to list AWS connections, then make a targeted request. If several accounts are connected, name the one to use.
Choose permissions in AWS
Creating the role and its trust policy grants the displayed d5s principal access under the attached IAM policy. Verification confirms the connection; it does not create the AWS trust or grant additional permissions.
New role templates default to Infrastructure diagnostics across All regions. This profile attaches AWS-managed SecurityAudit, ViewOnlyAccess, and CloudWatchReadOnlyAccess, plus supplemental diagnostic reads. It provides broad inventory, configuration, health, metrics and log visibility. Logs and configuration can contain sensitive data. AWS maintains and updates these managed policies.
Application read access uses AWS ReadOnlyAccess for broader reads, including supported application data. Operational access uses AWS PowerUserAccess and permits resource creation, changes and deletion; it is not full IAM administration. Review the selected policies before applying a template. Authorized agents can use the effective role permissions. These template choices do not update a role already installed in AWS.
Selected regions restricts regional API endpoints using an explicit deny. The listed global services remain accessible when permitted by the chosen profile: IAM, Organizations, Account, Route 53, CloudFront, Support, Budgets, Cost Explorer, Pricing, Health, Shield, WAF Classic, Network Manager and Route 53 Domains. This endpoint restriction is not a data-residency guarantee. Your d5s hosting region and the customer AWS regions are separate choices. The connection Region is the default for operations; the tool's region field selects another permitted Region.
IAM boundaries, organization policies, resource policies and explicit denies still apply. Keep existing roles under your infrastructure workflow; review any plan that changes their permissions. Roles created from an earlier template keep their original permissions.
Event-triggered tasks use the trigger creator’s current permissions and existing workspace AWS connections. They have the same AWS actions as a normal conversation, within the connected role’s IAM permissions. They do not require a scheduled-check grant.
An admin role permits agents to create, change or delete AWS resources. A role that allows Secrets Manager, SSM decrypted parameters or resource data may return those values to the agent. Choose the policy for the data and actions you want to delegate. Agents never receive the role's credentials. Intentionally requested API data can appear in the agent's answer and follow normal conversation retention.
An existing IAM user key pair can provision this role through your normal infrastructure workflow. d5s does not accept access-key uploads.
Use AWS API operations
Ask your agent to use the aws tool with structured input. Use action: "list" to find verified connections and starter examples. These examples are not the complete set of permitted operations. AWS IAM accepts or denies each operation.
Use action: "run" with a lowercase service, an operation in PascalCase, and a JSON parameters object. PascalCase joins words with an initial capital, as in DescribeInstances. Use the exact AWS API parameter names and JSON values. Omit parameters when the operation has no input. Use the connection_id returned by list when several accounts are connected.
For example, this input requests a monthly cost breakdown in the selected account:
{
"action": "run",
"service": "ce",
"operation": "GetCostAndUsage",
"region": "us-east-1",
"parameters": {
"TimePeriod": { "Start": "2026-09-01", "End": "2026-10-01" },
"Granularity": "MONTHLY",
"Metrics": ["UnblendedCost"],
"Filter": { "Dimensions": { "Key": "REGION", "Values": ["eu-central-1"] } },
"GroupBy": [{ "Type": "DIMENSION", "Key": "SERVICE" }]
}
}
The optional region field overrides the connection's default Region when IAM allows it. Use s3 for S3 API operations. CLI command strings, shorthand flags, local files, profiles, credentials, custom endpoints and event streams are not supported. The tool does not accept executable code.
d5s uses the managed AWS MCP endpoint in its hosting Region. EU hosting uses Frankfurt. US hosting uses Northern Virginia. This does not change the Region selected for your AWS API operation.
Output is JSON. Reads can automatically fetch multiple pages. Responses are bounded to 24,000 bytes and marked partial when trimmed. A partial result does not establish full inventory coverage. Narrow filters or time windows. A page-size parameter does not bound the total when AWS fetches more pages. Use any returned continuation token for further reads. Do not repeat a mutation to recover omitted output; inspect its state with a separate read.
d5s polls accepted tasks within a bounded request deadline. It does not automatically repeat an operation. A pending, timed-out or unconfirmed result can still complete in AWS. Inspect AWS state before submitting it again. Set the time window explicitly for log and cost investigations.
Tool history retains an operation receipt, without parameter values or raw AWS results. Requested data can still appear in the agent's answer under normal conversation retention. A request receipt identifies the tool invocation; it is not proof of a downstream AWS CloudTrail request ID.
Read PostgreSQL data
AWS infrastructure access does not grant SQL access to PostgreSQL. To query data or database statistics, use a separate PostgreSQL connection. Choose direct verified-TLS access or an on-demand SSH tunnel through a customer bastion. The bastion can reach a private RDS database without exposing the database publicly. Dedicated database permissions, connection sharing and tool approvals still apply. Saving the connection starts no background health checks.
Slack and Teams authorization
AWS workspace connections are shared capabilities of a coworker bound to a Slack or Teams channel. Channel members can ask that coworker to use its configured role under the existing conversation grant. A Slack DM uses its linked d5s user and current resource permissions. d5s checks the person's current access on each request. If access is removed, the request is refused. Audit records name the external person, not the agent owner.
Connecting a shared coworker to Slack or Teams delegates its configured shared capabilities to that conversation. If two audiences must not influence one another, use separate agents and appropriate IAM roles. Work started by another agent cannot use AWS access. Scheduled wakeups need the separate read-only grant described below.
Scheduled cost audits
A workspace connector admin can authorize one recurring, in-app agent wakeup without an end date or run-count limit to use one verified AWS connection. This is an explicit grant. The agent owner and the schedule itself do not grant AWS access.
Create the wakeup first. Then create a scheduled grant for that wakeup and connection with the user-authenticated API. The same API lists and revokes grants. Only a workspace connector admin can use it. See the generated OpenAPI reference described in the Developer guide for exact paths and fields.
The admin must retain workspace connector-admin permission and permission to run the agent. Scheduled reads stop if either permission is removed, or the wakeup, agent, or connection changes.
The grant covers the approved wakeup configuration. Editing its instructions, timing or delivery invalidates the grant. Authorize the new configuration before its next run. Pausing the wakeup, revoking its connection or deleting its grant stops its AWS reads. External delivery is not supported for these grants.
Scheduled operations can only read approved infrastructure metadata, ECR manifests and lifecycle policies, CloudWatch utilization, Cost Explorer, pricing and Savings Plans. The customer role must also permit each read. The grant permits no application data reads, log contents, image deletion, infrastructure change or commitment purchase.
For a cost report, separate usage, credits and net charges. Use complete comparison periods, record AWS data freshness, and page bounded inventories before claiming full coverage. Recommendations do not authorize cleanup or purchases.
Trusted email authorization
AWS access from email requires a d5s-managed mailbox, a verified sender domain, and the exact sender address in the mailbox's Trusted senders rules. A wildcard or domain-only trusted rule, correspondence history, a one-time quarantine release, an agent owner, or a matching From header is insufficient. The sender must still satisfy current inbound rules, the mailbox and contact must remain active, and the user who explicitly saved the trust policy must still be active with workspace membership and permission to run the agent.
This is an explicit delegation to that email address, not proof of an individual teammate's identity. Audits attribute the external sender and exact mailbox. AWS access from email is available only from d5s-managed mailboxes. From other mailboxes, use the web or a connected Slack or Teams conversation.
Revoke or troubleshoot
If the trusted d5s principal changes, update the role's trust policy through your infrastructure workflow, then verify the connection again. Use the exact principal and external ID shown in setup. Ordinary application updates do not require a trust-policy change. Deleting and recreating a trusted IAM role can break trust even when its name is unchanged; see AWS role principal guidance.
A workspace admin can revoke a connection under Cloud accounts → Details → Revoke access. Confirm the named account to revoke it. Remove the role trust or role in AWS as well to close the AWS side. Revocation blocks new operations and suppresses results from a call whose connection or source grant was revoked. It cannot undo a mutation already sent to AWS. d5s records operation attempts and outcomes without raw inputs or results, including attempts whose result could not be delivered. AWS CloudTrail supplies upstream request detail for the events it records.
If verification fails, check that the role account matches the connection and the trust policy uses the displayed regional d5s principal and exact external ID. When AWS rejects an operation with AccessDenied, AccessDeniedException or UnauthorizedOperation, the agent receives a permission error naming the failed operation and Region. Review the connection’s IAM permissions, account policies and Region restrictions. In web chat, Review permissions opens guidance for that existing connection. It identifies the denied operation, Region and required IAM action when available. Successful parts of the investigation remain valid; denied parts are reported as incomplete. Review the policy in AWS through your approved infrastructure workflow. d5s does not broaden IAM permissions. After your changes, continue the read-only investigation to check the denied operation. Role verification alone does not prove it is allowed. Scheduled read restrictions are separate from IAM permissions: widening IAM does not bypass them.
A permission rejection does not mean AWS is unavailable. A role-assumption rejection usually means the role's trust policy does not match the displayed d5s principal and external ID. Verification establishes the connection identity, not permission for every API.
If AWS rejects parameters, the error identifies the operation, Region and recognized AWS error code without repeating parameter values. Use the AWS API reference to check parameter names, required fields and types. For CloudWatch metrics, use JSON arrays such as "Statistics": ["Average", "Maximum"] or "ExtendedStatistics": ["p95"]. Correct the parameters before submitting the operation again. CloudWatch metric reads use the CloudWatch service, including metrics for RDS databases. A rejected request does not by itself prove an IAM denial.
If AWS reports that a requested resource does not exist, the agent receives a not-found answer. An example is an ECR repository with no lifecycle policy. The error identifies the operation, Region and AWS error code. This answer does not mean AWS is unavailable. Check the resource name and Region before you ask for the operation again. A resource created moments ago can take a few seconds to become visible.
A timeout or unavailable result leaves the resource state unconfirmed. It does not prove that a resource or lifecycle policy is absent. Keep successful findings and report the remaining coverage gap.
An AWS access connection is separate from Kubernetes Satellite. AWS public APIs use the account role; Kubernetes API and private VPC service access require their own installation and grant.