Skip to main content

Files, Skills, Packages, Apps and Accounts

Updated on 10 November 2026.

Besides sensitive-data rules (Policies and rules), the Controls tab governs file uploads to AI apps, skills and packages in coding tools, which applications run, and which AI accounts are allowed.

Go toQuilrAI ConsoleGovernPolicy EngineEndpoint AgentDesktopControls

Key terms​

  • AI app: an AI chat or assistant application, such as ChatGPT, Claude or Gemini, in a browser or desktop app.
  • Coding tool: an AI coding agent on the device, such as Claude Code, Cursor or Codex CLI.
  • Sensitivity label: a classification tag stored inside a document, such as "Confidential".
  • Skill: a folder of instructions a coding agent can run, often as a slash command such as /review.

Where to find these controls​

  1. Open Govern > Policy Engine and select Endpoint Agent in the workspace selector.
  2. Select Desktop, then the Controls tab.
  3. Choose an area from the left menu. Packages opens Settings > Endpoint Agent.

How each control goes live​

  • Files: saved to your draft. Devices use it only after Review & publish. See Draft, test, publish and restore.
  • Skills, Apps, Accounts and Packages: saved at once and applied when each device next downloads its settings. They have no draft and are not part of a policy version: Test does not evaluate them, History does not list them, and restore and Turn off do not change them. See How long changes take to reach devices.

While enforcement is paused, file rules in AI apps only monitor. Skills rules, Packages checks, file-type rules in coding tools and the Accounts checks that coding tools make themselves are off. Apps and other Accounts checks keep applying. See What changes while enforcement is paused.

Files​

A file rule checks the type or sensitivity label of a file that someone uploads or attaches to an AI app. It is a Policy Engine rule with a file condition, built and published like a sensitive-data rule, and applies only while the Policy Engine is on.

Files lists every rule with a file condition, across all policies. A row's Status switch turns the rule on or off in your draft.

note

File rules apply in AI apps only. In coding tools, file checks still come from File Restrictions in your previous Endpoint settings (Console V1), which match file type or sensitivity label. See What your previous Endpoint settings become.

Create a file rule​

  1. In Controls, select Files, then Create policy. The first rule already has a file condition.
  2. On Start from, keep Blank (a starter policy removes the file condition), continue, and enter a Name.
  3. On Who, choose Everyone, Only these people or Everyone except, then people, Smart Groups or departments.
  4. On Where, choose the AI apps. Coding tools is greyed out.
  5. On Looks for, set the condition under Files. Add a file condition adds another.
  6. On Response and risk, choose What happens in AI apps.
  7. Complete Priority and Check and save (Create a policy), select Create policy, then test and publish.

File conditions​

ConditionOperatorsHow it matches
File typeis one of, is not one ofRead from the file's contents, not its name. Plain-text files such as CSV, TXT or JSON may not be recognized.
Sensitivity labelis one of, is not one of, contains, does not containis one of compares whole labels, ignoring case. Not every AI app sends labels; the editor warns when one in scope doesn't.
  • is not one of and does not contain match only files with a detected type or label.
  • Older rules can show File kind, which still applies until removed.
  • When several rules match the same upload, see How policies are evaluated for which response wins.

What the device does​

ResponseUpload result
BlockStopped, with a block message.
Require a reasonHeld until the person gives a reason; blocked if none comes in time.
Ask for a reasonContinues, with a reason popup.
Record onlyContinues, with no message.

Each matched upload is recorded in Findings & Interactions. Without a decision in time, the upload continues.

Commands & tools​

Commands & tools is tagged Not available. To limit a Sensitive data rule to commands or MCP tools, use Narrow to specific actions in its Where step.

Skills​

Skill rules allow or block skills in Claude Code. The Endpoint Agent reports the skills installed on each device (discovered skills), and a rule matches only a discovered skill. Smart Groups are not available.

  • Rule for one person: blocks or allows a skill by its exact folder. Edit later changes only the response.
  • Rule for everyone: always blocks, by content or name, also for people who install the skill later. To change it, remove it.

If Add skill rule shows no Who choice, rules for everyone are not enabled; contact QuilrAI support.

When Claude Code runs a skill, either through its Skill tool or because the person types /name, any matching Block rule that covers the signed-in person stops it. Allow never overrides Block; skills with no matching rule run. Slash-command checks need Claude Code 2.1.116 or later.

Block or allow a skill for one person​

  1. In Controls, select Skills, then Add skill rule.
  2. If Who is shown, choose One person, then the Person.
  3. In Skill location, select Choose a discovered skill, or type the folder exactly as reported, such as ~/.claude/skills/review. Wildcards are not allowed.
  4. Optionally, enter a Skill name (display only).
  5. Under Skill response, choose Block or Allow, then select Create skill rule.

Block a skill for everyone​

  1. Select Add skill rule, then under Who choose Everyone.
  2. Under Match, choose This skill's content (that content in any folder, picked with Choose a discovered skill) or Any skill with this name (every skill with that exact name, including unrelated ones).
  3. Choose the Coding tool, then select Block for everyone.
warning

Devices enforce a rule for everyone only when it matches by This skill's content and its Coding tool is Claude Code or Any coding tool. Other rules for everyone, such as Any skill with this name rules, are saved but not enforced.

Removing a rule for everyone does not remove per-person rules for the same skill.

Packages​

Dependency scanning checks each package (software library) that Claude Code installs for known vulnerabilities, such as CVEs (publicly listed vulnerabilities), and newer versions. It is a setting, not a policy.

Go toQuilrAI ConsoleSettingsEndpoint Agent
Switch (all off by default)What it does
Claude Code dependency scanningRecords installed packages for Inventory. Alone, it does not check or stop installs.
Block vulnerable dependenciesStops installs of versions with known vulnerabilities, with a summary.
Notify about newer dependency versionsTells the developer a newer version exists. The install continues.

Scanning covers install commands and supported manifest changes for npm, PyPI, crates.io, Go, RubyGems, NuGet, Packagist and Maven. If a package can't be assessed or the check times out, the install is allowed.

Turn on dependency scanning​

  1. On the Dependency scanning card, turn on Claude Code dependency scanning.
  2. Turn on one or both response switches. They stay greyed out until scanning is on.
note

Dependency scanning is supported in Claude Code only. A row such as Cursor install checks means checks were turned on for that tool earlier. Select Turn off to clear them. Dependency inventory is not affected.

Apps​

An app rule blocks or allows an installed application for everyone, a Smart Group or one person. When a blocked application starts, the Endpoint Agent ends it and its child processes, but never operating-system components. If the person started it, they see a popup and a finding is recorded. If it starts again, the Endpoint Agent ends it again.

Block or allow an application​

  1. In Controls, select Apps, then Add app access rule.
  2. In Installed application, choose the application. It lists apps reported in the last 30 days; type to search.
  3. In Operating system, choose macOS and Windows, macOS or Windows.
  4. Under Applies to, choose Everyone, A Smart Group or One person (enter their Person's work email, as used on the device).
  5. Under What should happen?, choose Block app from running (default) or Allow app to run.
  6. Select Create app access rule.

Which rule wins​

  • A person's rule overrides their Smart Groups' rules, which override the rule for Everyone.
  • For the same audience, a rule for one operating system overrides one for macOS and Windows.

Notes​

  • Block (existing) is an earlier blocking rule, shown as Keep existing block when edited. It stops the app but does not isolate its files.
  • An app reported as several programs is one row with a parts tag. The rule, Edit and Remove cover every part.
  • Unidentified app rules name an app only by App identifier. They still apply and can be removed but not edited.
  • Audit-only app access is not available.
  • Edit changes only the response. For another application, audience or operating system, add a rule.
  • Inventory approval labels, such as Approved, do not stop applications.

Accounts​

Account restrictions decide which accounts, plans and models people may use in an AI app. They apply to everyone, without per-person exceptions. Each check allows only the values you enter.

Checks for each app​

Supported apps are ChatGPT (also Codex), Claude (also Claude Code), Microsoft 365 (including Copilot), Granola, Cursor, GitHub Copilot, Antigravity and Gemini. Any extra checks appear under Other restrictions.

CheckAppsWhat you enter
Approved workspacesChatGPT, Claude (Anthropic workspaces), GranolaWorkspace IDs. Claude's (wrkspc_...) are in the Claude Console under Settings > Workspaces.
Approved organizationsClaudeOrganization IDs (UUIDs), in Claude under Settings > Account > Organization ID.
Approved teamsCursorTeam IDs. Personal accounts have no team and are not blocked.
Allowed plansAll except Microsoft 365 and GeminiA listed plan, or a plan name or full plan ID.
Allowed modelsSame as Allowed plans. For Antigravity it can be saved but is not enforced; use Allowed plans.Model names. gpt-5* allows every model starting with gpt-5; * alone allows all.
Approved organizationsMicrosoft 365Domains (contoso.com) or tenant IDs allowed at Microsoft sign-in, plus Your company's Microsoft directory ID (a GUID).
Personal accounts and tenant policyMicrosoft 365Tenant Restrictions v2: Block personal Microsoft accounts, or Use a Cross-Tenant Access policy with <directory ID>:<tenant restrictions policy ID> from Microsoft Entra Cross-tenant access settings.
Work accounts onlyGeminiOn/off. Blocks personal Google accounts and signed-out use; other Google services are unaffected.
Approved domainsGeminiEmail domains, matched whole (example.com does not allow mail.example.com). Also turns on Work accounts only.

Restrict accounts for an app​

  1. In Controls, select Accounts, then Edit on the app's row. View means read-only access.
  2. Under Restrictions, turn on each check you need and enter its values, separated by commas or new lines.
  3. For ChatGPT or Claude, keep or clear Also enforce at OpenAI or Also enforce at Anthropic.
  4. Select Save account access.

A turned-off check keeps its values until you select Clear on its card and save.

Enforcing at OpenAI or Anthropic​

For ChatGPT and Claude, the approved-list card has an Also enforce at OpenAI or Also enforce at Anthropic checkbox:

  • Selected (default): the vendor also refuses other workspaces or organizations, including at sign-in. For ChatGPT, this needs ChatGPT Enterprise. For Claude, it also covers Claude Code, the API and the Claude Console, and people see Anthropic's "Access restricted by network policy" page.
  • Cleared: only the Endpoint Agent checks prompts, so people can still sign in and switch to an approved account.

What the device does​

A prompt from a disallowed account, plan or model is blocked with Workspace/Org ID not approved, Plan not approved or Model not approved, and a finding is recorded. See What users see and findings.

  • On claude.ai, if Anthropic refuses an organization, the agent switches the person to an approved one where possible, or signs them out.
  • For Microsoft 365, Microsoft Entra refuses a disallowed account's sign-in, which blocks that account in every Microsoft 365 app, not just Copilot.

Notes​

  • Only traffic the Endpoint Agent inspects is covered. Browser Extension policies are separate.
  • A check that is on with no values allows everyone (tagged On, nothing saved).
  • Checks tagged Needs a current agent are not enforced on older agents until they update.
  • For Microsoft 365, if both cards are on, both restrictions are sent at sign-in. Once Tenant Restrictions v2 is in place, Microsoft recommends turning that app's Approved organizations off.

Permissions​

AreaViewChange
FilesPolicy page and Read policiesPolicy page and Update policies
Skills, Apps, Accounts, PackagesEndpoint Agent page and Read Endpoint AgentManage Endpoint Agent. With V1 roles, Apps needs App Management (Tenant) Update.

With Read policies but not Update policies, a Needs Policy update access bar shows above every Controls area. You can still change Skills, Apps and Accounts if you have Manage Endpoint Agent. See Permissions.