> ## Documentation Index > Fetch the complete documentation index at: https://docs.growthbook.io/llms.txt > Use this file to discover all available pages before exploring further. # API Keys for Agents > Name the member an agent works for with X-GrowthBook-Requested-By, require a person to confirm risky changes before they run, and set when keys expire and rotate. Agents reach GrowthBook through the [REST API](https://docs.growthbook.io/app/api.md), with a Personal Access Token or a Secret Key (the [MCP server](https://docs.growthbook.io/integrations/mcp.md) and the [Agent Skills plugin](https://docs.growthbook.io/integrations/ai-agents/agent-skills.md) both work this way). Three key settings help keep that work accountable: * **X-GrowthBook-Requested-By** names the member who asked for each change, so history shows who it was for. * **Human-in-the-Loop** holds risky changes until a person confirms them in GrowthBook. * **Expiration** retires a key on a date, so access an agent no longer needs doesn't linger. Edit a Secret Key under **Settings → API Keys**, or your own token under **Personal Access Tokens** in your user menu. Requested-By applies to Secret Keys only, since a Personal Access Token already acts as its owner. Confirmation applies to both. ## Name who asked A request made with a Secret Key can send an `X-GrowthBook-Requested-By` header with a member's id or email address. History, revision logs and webhooks show that member next to the key. Each Secret Key acts as **System** or **User**: * **System** (the default): requests act as the key, with its own permissions. The header is optional and only records who asked. * **User**: every request must send the header, or it's rejected. The request acts as that member, with only what both the key and the member can do. The header isn't verified, so whoever holds the key can name any member. See [Naming who asked](https://docs.growthbook.io/api/introduction.md#naming-who-asked) in the API docs for the full rules. ## Permissions Choose what a key or token can do with a preset: * **Read only**: reads work, and every write is blocked. * **Human-in-the-Loop** (Enterprise): risky or destructive writes wait for a person to [confirm them](#require-confirmation). Not offered for a key that acts as System, since no one in particular is behind its requests. * **Full access**: everything the token's owner, or the member a key names, can do. A key that acts as System gets full admin access. Choose **Customize** to pick a base role, limit it by Project or environment, and set confirmation rules yourself. On a plan with only the admin role, **Full access** is the only option. The same choices appear on the consent page when an app such as the [MCP server](https://docs.growthbook.io/integrations/mcp.md) signs in with OAuth. ## Require confirmation With the **Human-in-the-Loop** preset, or **Require confirmation** on under **Customize**, a write the key's rules cover isn't run when the agent sends it. GrowthBook holds it and answers with a link to a confirmation page. A person opens the link, reviews the change, and confirms or rejects it, optionally leaving a note for the agent. A confirmed request then runs, with the key's permissions. The confirmation page shows: * **The change**: a summary, such as "Update Project Checkout: name", with links to what it affects * **The request**: its method, path and body * **Checks it skips**: anything the request asks to bypass, and what that check said when it last refused the agent ### Recommended or all actions New rules default to **recommended actions**, which hold: * Publishing, archiving and deleting * Starting, stopping or changing live Experiments and ramps, including their targeting and traffic * Skipping a check, such as `ignoreWarnings`, `skipHooks` or a required review, even when the change itself wouldn't be held * Editing settings, members and teams, environments, Projects, SDK Connections, custom hooks, attributes, custom fields, metrics, fact tables, segments and dimensions Creating things, working on drafts, AI Visual Editor changes, and writes like dashboards and comments run without confirmation. Arming a publish or an Experiment start to happen later, on approval or on a schedule, counts as the publish or start itself, including any archive or ramp plan it carries. Editing a draft that's armed counts as everything its publish could do, and editing an Experiment that's scheduled to start counts as the start plus the change as if it were live. If a key whose publishes need confirmation edits an armed draft any other way, GrowthBook cancels that draft's schedule or publish on approval, so nothing it changed goes live without a person. Creating a Feature Flag that's on in an environment counts as publishing there. If your organization lets REST requests bypass reviews, a key's publishes and archives of Feature Flags, Saved Groups, Constants and Configs also count as skipping a review. Choose **All actions** to hold every write, including creates and drafts. Reads and calculations never wait. To narrow or widen this, choose **Customize behavior** and add rules. Each rule picks actions and, optionally, environments, and a write is held when any rule matches it. A rule limited to some environments still holds actions that don't belong to an environment, such as editing settings. ### Who can confirm * **A Personal Access Token**: only its owner. * **A Secret Key that names a member** with X-GrowthBook-Requested-By: only that member. * **A Secret Key otherwise**: any member of the organization. A confirmed request runs as the key, not as the person who confirmed it. Anyone who can use an agent with the key can already act through it, so limit who can use a key to limit who can confirm its requests. Each decision is recorded in the audit log, with who made it. Confirming isn't a review approval. A change that needs approval still needs it. ### Limits * A request waits an hour for a decision, then expires. * A request that publishes, reverts or archives, or that changes a Feature Flag or an Experiment's status, schedule, targeting or traffic, isn't run if what it acts on changed after it was held. Other writes, such as editing settings or a draft, run as sent, so they can overwrite changes made while they waited. The agent can send a request that wasn't run again for a new confirmation. * A confirmed request runs once. If it fails, the agent gets the error and can ask again. * Confirmation rules apply to Secret Keys, Personal Access Tokens, and OAuth sign-ins such as the hosted MCP endpoint, which choose them on the consent page. To build an agent that handles held requests, see [Waiting for confirmation](https://docs.growthbook.io/api/introduction.md#waiting-for-confirmation) in the API docs. ## Expiration When you create a key, choose when it expires: after 1, 7, 30, 60, 90, 180 or 365 days, on a date you pick, or never. You can change it later by editing the key. To replace a key's secret, [rotate it](#rotate-a-key). To start a different key from an existing one's settings, choose **Copy settings to new key** from its menu. Before a key stops working: * Every API response carries its expiration in the `X-GrowthBook-Key-Expires-At` header. A request made with a rotated key's previous secret also carries `X-GrowthBook-Secret-Expires-At`, when that secret stops working. * GrowthBook shows a notice when a key expires within 7 days, or has expired while something still uses it. You see your own Personal Access Tokens, and members who can create or edit Secret Keys see the organization's. * Secret Keys trigger the `apiKey.expiring` [webhook event](https://docs.growthbook.io/app/webhooks/event-webhooks/events.md) a week before they expire, and `apiKey.expired` when they do. A request made with an expired key fails with a `401` that says where to create a replacement. One made with a previous secret after its grace period ends, or after it's revoked, fails with a `401` that says the key was rotated and where to reveal the new secret. ### Rotate a key Choose **Rotate key** from a key's menu to give it a new secret. Its id, role, permissions and settings stay the same, so nothing tied to the key needs to be set up again. * Choose when the current secret stops working: immediately, or after 1 hour, 24 hours or 7 days. The dialog shows the exact times before you confirm, and the new secret once you do. * Rotating restarts the key's expiration: it keeps the lifetime it had, counted from now, within your organization's maximum. * While the previous secret still works, the key is marked **Previous secret active**, with when that secret was last used, so you can tell when every client has switched. **Revoke previous secret** stops it early, for example if it leaked. * Only one previous secret is kept, so rotating again stops the earlier one right away. * An expired key can't be rotated. Create a new one instead. Rotating a Secret Key takes permission to create API keys, and only its owner can rotate a Personal Access Token. Each rotation is recorded in the audit log. ### Organization policies Admins can require keys to expire within a maximum lifetime: under **Settings → API Keys** for Secret Keys, and under **Settings → Manage PATs** for Personal Access Tokens. Turn on **Require secret API keys to expire** (or **Require personal access tokens to expire**) and set the **Maximum lifetime (days)**. * New keys must expire within the maximum, and **No expiration** is no longer offered. * The maximum counts from when a key was created or last rotated, so editing a key can't push it past the limit. Shortening is always allowed. * Existing keys aren't changed. Ones with no expiration, or a later one, are counted as non-compliant on the policy and marked in the **Expires** column. Click the count to have them expire the maximum lifetime from today.