Inference hooks — a feature that screens requests to Claude before they're answered — lets Claude Enterprise organizations inspect prompts, tool call responses, and uploaded file text on their own server, then allow or deny each one before Claude responds. This guide summarizes what Inference hooks are, who can use them, and how they work, based on Claude's official Help Center documentation.
🟢 Model references match the current lineup · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (
What are Inference hooks?
Inference hooks let Claude inspect every prompt, tool call response, and uploaded file text before Claude starts generating a response. They're built for a compliance team to check requests against its own policy, and they're currently available in beta (a pre-release, testing version) on Claude Enterprise plans.
Who can turn them on?
Only an organization's Owners and Primary Owners can turn on and manage Inference hooks. Regular team members or accounts without admin permissions can't access this setting, so if you want to use it, first find out who in your organization holds one of these roles.
How do Inference hooks work?
Once turned on, Claude sends every prompt to a server your organization hosts before it starts generating a response. That server checks the prompt against your policy and answers allow or deny, and Claude only continues once it has that answer. Because this check happens inside Claude's infrastructure rather than on someone's device, it doesn't depend on anything being installed on employees' computers.
What does it cover?
One setup covers your whole organization — Claude, Claude Code, and Cowork, plus every other Claude Enterprise product — including tool calls made through skills, plugins, and connected tools. You don't need to configure each product separately.
What gets checked?
Inference hooks inspect three kinds of content.
| Checked content | Description |
|---|---|
| Prompt | The request text a user types directly to Claude |
| Tool call response | Results Claude receives back through skills, plugins, or connected tools |
| Uploaded file text | Text contained in a file a user attaches |
Why would you use it?
Common uses include data loss prevention (DLP — stopping sensitive information from leaving your organization), real-time transcript archival, and enforcing your own organization's policies. It's especially useful for organizations in regulated industries that want to apply compliance controls consistently across all Claude usage.
What you need before you start
- An active Claude Enterprise plan
- Owner or Primary Owner permissions in your organization
- Development resources to build and host your own policy server (self-hosting is required)
How to turn it on (overview)
- Confirm eligibility — Verify your organization is on an Enterprise plan and that you have Owner or Primary Owner permissions. Success looks like: a security/compliance settings area is visible in the admin console (confirm the exact menu label on screen).
- Build your policy server — Implement your policy endpoint, including request signature verification. Follow the official "Inference hooks" technical documentation for implementation details.
- Turn it on in the admin console — An Owner or Primary Owner enables the feature in the admin console. Confirm the exact menu location and label on screen.
- Verify it's working — Send a test prompt to confirm your policy server returns allow/deny responses correctly. Success looks like: allowed requests get a normal Claude response, and denied requests are blocked.
Note: Claude only continues after receiving an allow/deny answer from your policy server. If your server doesn't respond, Claude can't proceed — so monitor your server's uptime and response time closely.
Common issues & fixes
- Because this is a beta feature, the relevant menu may not appear for every organization → check with your admin whether your plan and the beta feature are both active.
- Only Owners and Primary Owners can access the setting → ask an admin with one of these roles to configure it.
- If your policy server doesn't respond, requests can be left waiting → check your server's availability and review logs with your technical team if needed.
Next steps
Inference hooks apply across Claude Enterprise products, including Cowork. Claude's Help Center also covers related topics such as "Use Claude Cowork on Team and Enterprise plans," "Manage custom roles on Enterprise plans," "Monitor Claude Cowork activity with OpenTelemetry," "Claude Cowork architecture overview," and "Enable US-only inference for your organization" — refer to those articles along with the official technical documentation for exact setup steps.
Frequently asked questions
Q. Can anyone turn on Inference hooks?
No. Only Owners and Primary Owners of a Claude Enterprise organization can turn on and manage this feature.
Q. Does Claude provide the policy-checking server?
No. Your organization hosts and runs the policy server yourself. Claude sends prompts to it and waits for an allow/deny answer before continuing.
Q. Do I need to configure Claude, Claude Code, and Cowork separately?
No. One setup applies to Claude, Claude Code, Cowork, and all other Claude Enterprise products, including tool calls made through skills, plugins, and connected tools.
Q. Is this a finished, generally available feature?
Inference hooks are currently available in beta on Enterprise plans, so availability and details may vary by organization.
Related articles
- How to Write Prompts for Claude: 6 Core Principles from the Official Docs
- What Is Claude Cowork? Claude's File-Working Desktop Mode
- What Is Claude Design? Anthropic's Chat-to-Canvas Design Tool
- Claude Memory — How to Make Claude Remember and Manage It
- Claude Styles — Lock Tone, Format, and Length Your Way