Cloud monitoring with read-only access lets a monitoring service retrieve permitted information without authorising it to change the cloud resources being monitored. Before connecting an account, check the actual permissions, which accounts and resources they cover, what data leaves the cloud provider and how you can revoke access. Read-only is a useful boundary, but it does not mean the service sees nothing sensitive or that connecting it needs no review.
For an agency, the decision also involves client authority. Permission to maintain a website should not be treated as automatic approval to connect a third-party service to the client's wider infrastructure. Establish who owns the account and who can approve the proposed visibility before starting the connection.
Read-only describes actions, not the whole boundary
A monitoring connection typically needs to discover resources, read their status and retrieve measurements. Depending on its purpose and permissions, it might also retrieve configuration details, resource tags and billing information. Those are different kinds of access, even when none permits a change.
Separate the assessment into three questions: who is connecting, what actions can that identity perform, and where can it perform them? The connecting identity might be a dedicated application identity or a role used by the monitoring service. Its permissions define the actions, while its scope defines the accounts or resources within reach.
Do not assume that seeing a database listed means the service can read its customer records. Reading database configuration is distinct from querying its contents. Equally, do not assume the word monitoring excludes access to logs or stored content. Ask for the permission list and the data collected, rather than inferring either from the dashboard.
Inspect the permission request before approving it
Ask for the policy, role assignment or token scopes used by the integration. These are the instructions that grant access. A statement that a connection is secure is not a substitute for knowing what it authorises.
- Match access to a purpose. Resource discovery, workload charts and billing reports should each have an explained reason for the permissions they require.
- Look for change permissions. Creating, updating, starting, stopping or deleting resources needs separate justification and does not belong in a strictly read-only monitoring connection.
- Check the boundary. Establish whether access covers one resource group, a subscription, an account or a wider collection of accounts. Ask whether that boundary can be narrowed.
- Check future coverage. Ask whether newly created resources inside the approved boundary become visible automatically, including resources unrelated to the original website.
Broad access is not automatically unjustified. A service discovering a changing estate may need wider visibility than one watching a fixed server. The decision is whether that reach matches the job and the account owner's expectations. If a client account contains unrelated business systems, account-wide visibility deserves explicit discussion.
If the policy is difficult to interpret, have someone familiar with the provider's access controls review it. Avoid solving a failed connection by granting administrator access. Identify the failed read operation and its purpose first.
Decide which information is acceptable to share
Read-only access reduces the risk of an integration changing infrastructure. It does not remove confidentiality risk. Resource names can reveal client names or unreleased projects. Tags may identify internal owners or cost centres. IP addresses and configuration details describe infrastructure, while billing data reveals spending.
For example, a shared agency cloud account might contain servers for several clients. A connection intended to investigate one site's workload could expose resource names and spending for other clients if its permissions cover the whole account. This is a hypothetical scope problem, not evidence that the monitoring tool is malfunctioning.
Ask the supplier which fields it retrieves, which it retains and who can view them in its own dashboard. Include exports, reports and support access in that discussion. Also ask how long collected information remains after disconnection. Removing access at the provider does not itself delete information already copied into another system.
Review credentials and the setup command
Use a dedicated integration identity rather than an employee's everyday login. This makes the connection easier to identify, review and remove without disrupting that person's other work. Where supported, prefer an approach that avoids handing over a long-lived secret, while still reviewing the trust relationship and permissions.
A read-only monitoring identity may still require an administrator to create it. These are separate permission boundaries: the person installing the connection can make access changes, while the resulting integration should receive only its intended rights. Review any supplied script or infrastructure template before running it with elevated privileges.
Check which identity the command creates, what permissions it assigns and what information it prints. If setup produces a secret, transfer it only through the intended secure connection flow. Do not put it in a support ticket, project note or screenshot. Keep a record of the identity and owner, not a copy of the secret in the access register.
Use a controlled trial to check coverage and removal
Start with a representative, limited environment where practical. A successful sign-in establishes that authentication worked. It does not establish that every intended resource or measurement is available. Conversely, an empty chart does not prove that a server is idle.
- Compare discovery with the provider's inventory. Confirm that expected resources appear and investigate anything outside the approved scope.
- Check freshness and gaps. Look for collection timestamps, access errors and unavailable metrics. Separate missing data from a genuine zero reading.
- Confirm collection costs. Ask whether the integration makes chargeable requests and how request volume is controlled. Read-only describes authority, not price.
- Test disconnection. Record the provider-side removal procedure, revoke the trial connection and check that fresh collection stops. Historical charts may remain visible, so their presence alone does not prove continued access.
Ask how revocation affects credentials or sessions already issued, and what delay to expect before collection stops. Removing a dashboard entry and removing provider-side access are not necessarily the same operation. Your runbook should distinguish them.
A concrete example of provider-specific access
TLDTrack's cloud monitoring connects to AWS, Azure and DigitalOcean with read-only access. Its documented setup uses a different access mechanism for each provider. That is why the review should follow the actual connection method, rather than a single generic approval.
- For AWS, the documented CloudFormation setup creates a role limited to read actions, with an external ID. TLDTrack states that no keys are stored and deleting the stack removes access.
- For Azure, the documented Cloud Shell command creates a sign-in with the Reader role on the subscription. The connection details are stored encrypted, and deleting the app registration removes access.
- For DigitalOcean, the documented connection uses a token with read scopes for account, Droplets, databases, monitoring and billing. The token is stored encrypted and is not shown again.
For the token-based connection, include provider-side token revocation in your review and trial. Do not assume that hiding a saved token or removing a screen from view invalidates the credential.
Leave an access record the next person can use
Before moving into routine monitoring, record the account owner, approver, integration identity, permission scope, purpose and revocation procedure. Name both the person responsible for access and the person handling monitoring failures. An expired or revoked credential should prompt investigation, not leave an apparently healthy dashboard displaying stale information.
Review the connection when client responsibilities change, infrastructure moves or the service is no longer needed. Keep repair authority separate from observation: a read-only monitoring alert still needs someone authorised to act in the provider account.
Approve the connection when you can explain what it reads, why that scope is justified, who can see the collected information and how to stop further collection. If any answer is missing, resolve that specific gap before connecting the production account.
