# Credential and integration hygiene > Credential age and rotation, stale or failing connections, orphaned servers, and integrations nobody uses any more. Source: https://triagic.com/checkups/credential-hygiene ## Scope [#scope] A sweep of the org's integration estate: what is connected, whether it still works, whether anyone still uses it, and what each connection would cost the org if its credential leaked. Runs on the connection inventory available to this run plus each provider's read-only identity calls. The subject is each credential the org holds for an external system — never this application's own configuration. **Never read, print, echo, or reconstruct a credential value.** Secrets are stored encrypted and no read-only tool exposes them; if any tool result contains something that looks like a key, token, password or connection string, that is itself a critical finding — report its location and shape ("a 40-character token in the description field of the X server"), never the value. ## Procedure [#procedure] 1. **Inventory the connections.** Using the connected-server data available to you, list every MCP server configured for this org: provider, display name, who created it, when it was created and last updated, whether it is shared at org level or personal to one user, and its current running state and last error. Personal servers matter here: a server owned by one user's account is a single-point dependency and a deprovisioning hazard when that person leaves. 2. **Health.** Group the inventory by state. Anything not running is either broken or abandoned, and the two need different actions: read the last error text. An authentication error means an expired, rotated or revoked credential; a network or spawn error usually means a host or runtime change. A server that has been failing for weeks with nobody noticing is a monitoring finding as much as a credential one — say so. 3. **Age and rotation.** Use the created/updated timestamps as the proxy for credential age: a connection whose credentials have not been updated in over a year is overdue on any rotation policy, and one over two years is a finding on its own. Where the provider exposes token metadata read-only, use the better source instead: GitHub and GitLab expose token scopes and expiry, Snowflake exposes `LAST_SUCCESS_LOGIN` and key-pair versus password authentication, cloud providers expose access-key age. Prefer provider truth over the local timestamp and say which one each finding used. 4. **Use.** Cross-reference the inventory against what actually runs: which connections real investigations and reports have drawn evidence from, using the ticket and investigation history available to you. A connected integration no work has touched is an *orphan* — it holds a live credential, widens the blast radius, and buys nothing. That is this checkup's highest-yield finding and the easiest remediation in the whole library. 5. **Blast radius.** For each connection, state what the credential can reach if it leaks: the account or project, and how broad the granted role is. A read-only role scoped to one schema and an account-admin role are the same "connected integration" on the card and wildly different exposures. Where a provider's read-only identity call exists (whoami, current-user, current-role), use it to confirm what the credential actually is rather than what someone intended it to be — the gap between those two is a common and unglamorous finding. 6. **Duplicates and drift.** Look for several connections to the same underlying system, particularly under different owners. Duplicates multiply the credentials in circulation, and usually one of them is the stale one nobody rotates. Note naming that hides which environment a connection points at (a server called "postgres" that is production). 7. **Deprovisioning check.** Cross-reference personal server owners against the current member roster: a server owned by a user who is no longer active is both an orphaned credential and a broken dependency waiting to surface. ## Finding keys [#finding-keys] The `key` names the *connection and the problem*, not this run, so a stale credential still stale next week updates one ledger row instead of opening a second, and a rotated credential that goes stale again is caught as a regression. Use `server::`, using the connection's stable key or name rather than a numeric id that could change. Never put dates or age values in the key. * `server:snowflake-prod:credential-stale` * `server:github-main:auth-failing` * `server:zendesk-legacy:orphaned` * `server:postgres-analytics:overbroad-role` * `server:mongo-personal:owner-inactive` * `secret-exposure:server-description:token-in-plaintext` ## Severity rubric [#severity-rubric] * **critical** — a credential value visible anywhere it should not be (a description, a playbook body, a ticket, a tool result); a live connection owned by a person who has left the org; a connection whose identity call proves it holds account-admin rights it was never meant to have. * **high** — a credential over two years old on a system holding customer or production data; a connection failing authentication for weeks with nobody noticing; an orphaned connection to a production system. * **medium** — a credential over a year old; an orphaned connection to a non- production system; duplicate connections to the same system under different owners; a personal-scope server that the whole team depends on. * **low** — naming and tidiness issues, an unused connection to a sandbox, a connection that has been failing since it was created and clearly never worked. * **info** — the estate summary itself: how many connections, by provider and state, recorded so next week's run can see what changed. ## Output guidance [#output-guidance] Open with a one-paragraph executive summary: how many integrations are connected, how many are healthy, how many are orphaned, and the single biggest exposure. Then sections for Health, Credential age and rotation, Orphaned and unused, and Blast radius, each naming where its evidence came from (the connection inventory versus a provider identity call). Close with a recommendations table (Connection | Issue | Recommended action | Urgency), rotations and disconnections first. Never print a credential value anywhere in the report — identify a credential by its connection, provider and owner only. If a secret was found somewhere it should not be, describe the location and shape and recommend rotation, not the value.