AI library

Source · docs

Inference hooks (Claude docs)

Author
Anthropic
Published
20 October 2026
Added
21 October 2026
Original
https://platform.claude.com/docs/en/manage-claude/inference-hooks

Summary

Inference hooks came out of beta on 2026-10-20 for Claude Enterprise organizations, according to the Claude platform release notes. A hook sends each governed prompt, and optionally each batch of tool calls, to an HTTPS "AI security server" that the organization runs. Inference continues only if the server answers allow. The hook runs on Anthropic's servers, so it covers claude.ai, Cowork, Claude Code, Claude Tag, Claude in Chrome, and the other Claude apps without installing anything on users' devices. The same release notes add user profiles and organization webhooks to the API.

Key points

  • Who can use it: Claude Enterprise only. It is configured by the Owner or Primary owner (organization:manage). It does not cover API (Platform) organizations, Amazon Bedrock, or Google Cloud.
  • Events: prompt fires before each governed inference request. tool_call fires before Claude's tool calls run, if Validate tool calls is on, and one verdict covers all the calls in a response.
  • What it sends: an HTTPS POST with the transcript, tool calls and results, and text extracted from attachments, signed under the Standard Webhooks specification. It never sends system prompts, tool definitions, or raw image or file bytes, so screenshots of documents are not inspected.
  • Verdicts: allow or deny only, within a timeout that defaults to 5 seconds. Hooks cannot redact or rewrite content. A deny shows the server's deny_reason plus an admin message, and is logged in the Activity Feed.
  • Rollout controls: shadow mode, a rollout percentage, role exclusions, a choice between failing open and failing closed, and a circuit breaker for a failing server.
  • After a deny: later requests resend the whole conversation, so the user must remove the content. In Claude Code, /rewind to the prompt that brought it in, or /clear.
  • Uses: inline DLP, live transcript archiving (always allow), prompt telemetry, and policy engines (model allowlists, working hours).
  • Also in the 2026-10-20 release notes: the /v1/user_profiles endpoints let a platform or reseller tag each API request with its end customer. A safety enforcement action can then target that customer instead of the whole organization. Admin API organization webhooks (beta header webhooks-2026-10-07, access through the account team) report when a user profile is restricted or banned, and when that ends.

Why it matters

Hooks give enterprises one server-side control point across every Claude surface, including Claude Code, instead of per-device agents. For the owner's own use, the practical points are that Claude Code sessions in an enterprise may be blocked by policy, and that /rewind is how to recover.