Skip to main content
Connect your AI

Endpoint agent policy

The policy tells every agent of an organization what to do. You set it on the Endpoint agent page in AI Governance, in the Policy section. An Admin can change it. An Auditor can read it and change…

Written By Christopher Scaminaci

Last updated 4 days ago

The policy tells every agent of an organization what to do. You set it on the Endpoint agent page in AI Governance, in the Policy section. An Admin can change it. An Auditor can read it and change nothing. If you manage client organizations, pick the Organization at the top first, because every choice below belongs to the organization you pick.

At a glance

  • What it does: sets what the agent does: allow, warn or block, the paste check and local AI tool control.
  • Plan: every plan finds AI. Blocking, the paste check and local AI tool control need the Pro plan or higher.
  • Who: an Administrator.
  • Where: AI Governance → Staff computers → Endpoint agent.

Agents read the policy about once a minute, and send their reports about every five minutes. A change reaches a computer within a few minutes. The agent keeps a copy of the policy on the computer, so a short outage at StackJack changes nothing.

What a plan includes

Finding AI on a computer and keeping your managed settings in place work on every plan. Blocking, the paste check and local AI tool control need the Pro plan or higher. Below Pro the policy forces every control to Allow, whatever you chose. The agent then finds and reports and enforces nothing.

The two switches

The two switches are in the On or off panel of the Overview tab, not in the Policy tab.

  • Use the endpoint agent. When it is off, installed agents stop acting on their next policy read. Use it as the first stop if something goes wrong.
  • Block AI tools when StackJack cannot be reached. The agent keeps working when StackJack is slow. It lets the action through and records that it was not checked. This switch only decides what happens when the agent's copy of your policy is older than the staleness limit on the AI gateway page. With the switch on, AI tools are blocked in that case. With it off, they stay open. The default is off.

Local AI tools

For each tool, choose Off, Watch with a hook or Wrap its MCP servers. Every tool starts at Off.

ToolHow the agent watches it
Claude CodeA hook, in the machine settings that StackJack makes.
CursorA hook, in Cursor's machine-wide hooks file.
CodexA wrap, in the staff member's own Codex settings.
Claude DesktopA wrap, in the staff member's own Claude Desktop settings.
Gemini CLIA wrap, in the machine settings that StackJack makes. Gemini CLI is wrapped and not hooked in this release, whatever you pick on the page.

Edit staff members' own AI tool settings is the switch for the two tools that live in a person's own settings file. It is on by default. The agent keeps a backup of the original file and can put it back. With the switch off, Codex and Claude Desktop cannot be wrapped. The other tools are not affected. Codex is also left alone when your approved MCP servers list allows only approved servers, because Codex checks a server by its command name and a wrapped server has the agent's name. How this works is in Local AI tool control.

The paste check

Check text pasted into AI apps turns the paste check on. It is on by default. When a person pastes into an AI app, the agent checks the text with the same data-loss rules as everywhere else in AI Governance. What happens follows your rules: the paste is allowed, the sensitive part is masked, the person is asked first, or the paste is stopped. The check works only when the app in front is an AI app from the agent's catalog. A paste through a right-click menu or by drag and drop is not checked. At the facts-only capture level no text leaves the computer, so the agent records that a paste happened and nothing more. See Capture levels and what your employer sees.

Unapproved AI apps and sites

Both settings have the same three choices. Both start at Allow.

ChoiceAI apps you have not approvedAI sites you have not approved
AllowThe agent reports the app and does nothing else.The agent reports the lookup and does nothing else.
WarnThe person sees a notice. The app keeps running.The person sees a notice. The site still opens.
BlockThe agent ends the app within about a second of opening.The agent stops the lookup of the site on that computer.

The AI apps and AI sites the agent knows come from a catalog that StackJack keeps up to date. The agent fetches it about once an hour, so a new entry needs no new agent version.

AI site control has a limit you should know. It works through a clearly marked section of the computer's hosts file, so it stops a lookup by name. It is not a filter on the traffic, and it does not open encrypted traffic. The agent writes only its own section and renews it regularly. It removes the section when the policy no longer blocks the site, and at uninstall. If the agent stops for good, a small cleanup task removes an expired section, so a crashed or removed agent leaves no block for longer than about 45 minutes.

When a security product blocks the same thing, you do not need both. See below.

Use my security product

For each of the three jobs (the paste check, AI app control and AI site control) you can leave the job to the agent, hand it to your security product, or let StackJack decide. The choice and the reasons are in Use your security product with the endpoint agent. Handing a job over is only a presentation choice. It never gives the agent or StackJack more or less access to your data.

Updates

  • How agents update. Self-update keeps each agent on the version StackJack names, and every new version is checked for its signature before it runs. RMM only means the agents never update themselves, and you update them through your own RMM.
  • Hold agents at version. Leave it empty to follow the released version. A version here pins every agent of the organization to it, also when that version is lower than the one an agent runs.

Enrollment keys

An enrollment key lets a device ask to join your organization. It grants no access to your data. The page shows a key once, when you make it. Rotate makes a new key, and the old key keeps working for 7 days so installs in progress can finish. To rotate you type ROTATE. Revoke stops a key at once. To revoke you type REVOKE. Devices that already enrolled keep working in both cases.

What a leaked key allows. Whoever holds a live key can enroll a computer into your organization. That alone gives no access to your data. If the person gives the same machine identity as a computer that already enrolled, the new enrollment takes over that computer's entry, and the real agent stops working at its next call to StackJack. To end this, rotate the key and, when it must stop at once, revoke it. Then look in the Devices list for a computer you do not know. A computer that was taken over this way sets its credential aside after about an hour and waits for a key, so enroll it again with a current key.

The uninstall code

With a code set, stopping or removing the agent on a computer needs the code. You can type a code (12 to 64 characters), have the page Generate a code, or Remove it. StackJack keeps only a check value, so it cannot show the code again. Write it down where your technicians can find it. With no code set, anyone with administrator rights on the computer can remove the agent. How to use the code is in Installing the endpoint agent.

Devices

The Devices list shows each computer that enrolled, with its platform and its health.

  • Healthy. The device reports.
  • Stale. The device has not reported in its window.
  • Tamper suspected. The agent was stopped without the uninstall code.
  • Revoked. You ended the device.

A revoked device stops working at its next policy read, and its machine cannot enroll again until you forget it. To revoke you type REVOKE. To Forget a device, which lets the same machine enroll again, you type FORGET.

A forgotten device still holds its old credential. When StackJack has refused that credential three times in a row over about an hour, the agent sets the credential aside and goes back to waiting for an enrollment key. Then install the agent again with a current enrollment key, and the computer enrolls again. Until you do, it shows no new reports.

A stopped or removed agent also raises an alert for your organization, once a day, with the number of devices.