Protect Claude Enterprise with inference hooks

Point your Claude Enterprise organization's inference hooks at WASViking so every request from claude.ai, Claude Desktop, the mobile apps, Claude Code and Cowork is inspected before the model runs, with nothing installed on devices and no traffic interception. Includes what each Claude plan, Team or Enterprise, allows.

Claude Enterprise Protection in WASViking^® AI Guardian answers your Claude Enterprise organization's inference hooks. Claude Enterprise sends every governed request, from claude.ai on the web, Claude Desktop, the mobile apps, Claude Code and Cowork, to one WASViking address before the model runs, and WASViking answers allow or deny. A denied request never reaches the model, and the person reads why.

Nothing is installed on devices and no traffic is intercepted: the vendor delivers the conversation itself, signed. This is the one path that inspects conversations in the Claude Desktop app, and it needs Claude Enterprise. The next section shows what each Claude plan allows; see also Which AI applications can go through the AI Gateway.

Claude Team or Claude Enterprise: what you can protect

Your Claude plan decides how far WASViking can see into Claude. Inference hooks are a feature the vendor offers only to Claude Enterprise organizations. They run on the vendor's servers, after a request leaves the person's device and before the model runs, and they are what lets WASViking inspect the desktop and mobile apps with nothing installed.

Claude Team has no inference hooks, and its desktop and mobile apps have no setting that sends requests anywhere else: the session goes straight to the vendor's service. No gateway, proxy or security product, from any supplier, can read or block those conversations. This is how the vendor designed the product, and it is the same for every customer. On Team, WASViking controls the app on company devices and inspects content where the vendor allows it, in the browser and in Claude Code.

If you are on Claude Team and need the conversations people have in Claude reviewed by your policies:

• Point people to claude.ai in the managed browser, where the extension applies your policies, and allow Claude Desktop only with the company account, or stop it from opening.
• Keep developers on Claude Code under Assistant Guard or the AI Gateway.
• Move the organization to Claude Enterprise and follow this page. Nothing changes on the devices.
• Or run Claude Desktop in the vendor's third-party inference mode with the company API key through the AI Gateway. Billing moves from seats to tokens; see Which AI applications can go through the AI Gateway.

The limits of the hooks themselves, which apply on Enterprise, are in the next section.

What it does and does not do

• The conversation has already reached the vendor when WASViking inspects it. A denial stops the model from processing it, not the transmission, and the denied message stays in the person's conversation. The person edits it, or starts a new chat.
• The answer is allow or deny only. Nothing is rewritten or masked on this channel.
• The whole conversation is inspected on every message, so a secret left in an earlier message keeps the conversation denied until it is edited. Repeats of the same finding are grouped in the activity.
• Attachments arrive as extracted text. An image without text is counted, not inspected. Tool calls are inspected before they run, and their results on the next message.
• The vendor does not say whether a request came from the desktop app or the web, and personal accounts outside the organization are not covered.
• When WASViking cannot answer, it answers an error, never an allow, and your organization's own failure handling at the vendor decides (block or allow). Those requests appear as "not answered", never as safe.

Before you start

• A Claude Enterprise organization. Only an Owner or Primary owner can configure inference hooks there.
• The Assistant Guard add-on in your WASViking contract, and the ai_guardian.gateway.manage capability (Admin or Manager).
• Rules on the coding assistant surfaces. The Standard preset in Assistant Guard is enough to start: it blocks secrets in prompts.

Step 1: turn it on and create a hook

In AI Guard, Protection, Claude Enterprise Protection, open the Configure tab. Step 1 holds the organization switch: turn it on and select Save. In step 2 give the hook a name and select Create hook. Copy the token: it is shown once. The Overview tab keeps a readiness checklist of what is still missing before production, with the state of the hook (off, connecting, observing or enforcing) computed from the real configuration.

Step 2: configure the vendor

Step 3 of the Configure tab, Connect WASViking to Claude Enterprise, shows the address and the header. On claude.ai, under Organization settings, Data and privacy, Inference hooks:

Step 3: observe, then enforce

While the hook observes, the Activity tab records what would be denied without denying anything. Review it with your rules, then choose Enforce policies in step 4 of the Configure tab, select Save mode and turn on Enforce verdicts at the vendor. Step 5 simulates a prompt with a secret and a tool call through the real decision engine before anyone is blocked. Changes at the vendor take about a minute.

Step 4: follow the activity

The Activity tab shows every request: the person, the application as the vendor sent it, the verdict, the rule, what was found and where (the latest message, an earlier one, an attachment, a tool input), and the verdict time. Filters: hook, event, verdict, would be denied, not answered, repeats, person. The People tab lists who was protected and who in the directory was not seen. A row opens the event details (metadata and masked evidence, never the content); Export CSV downloads the filtered events.

Optional: store a Compliance Access Key (read scope) in step 6, Advanced settings. WASViking then joins the vendor's own record of each denial to the event here, and shows when the vendor let requests through uninspected or tripped its circuit breaker.

Rotate the secret or the token

• Secret rotated at the vendor: paste the new one in WASViking. The previous one stays accepted for ten minutes, so deliveries in flight are not refused.
• Token: select Rotate token in step 2 and update the custom header at the vendor right after.

Troubleshooting

What next

• Protect Claude Code on Claude seats
• Protect coding assistants with Assistant Guard
• Which AI applications can go through the AI Gateway