September 20, 2026: Anthropic has expanded its Compliance API to include Claude in Chrome compliance transcripts. Announced on September 18, the addition is in beta for Claude Enterprise organizations and uses the existing Compliance Access Key with the read:compliance_user_data scope. Source: Claude Platform release notes.
What changed
The local-session endpoints now include conversations from the Chrome extension’s built-in chat, identified by product_surface: claude_in_chrome. Existing integrations can use that value to distinguish browser sessions from other supported Claude products.
The official session-transcript guide documents read-only endpoints for listing sessions, retrieving metadata and fetching messages. Transcripts include prompts, assistant responses, and tool calls and results. An Admin API key cannot substitute for the Compliance Access Key.
What the records can—and cannot—show
Anthropic says local capture applies while users are signed in with their Enterprise account and the organization has the Compliance API enabled. Records reflect requests reaching the API; they do not independently record everything that happened on the device. New sessions can take minutes to appear, and tool content is subject to truncation.
Why this matters for QA engineers
Our QA assessment: these records could help reconstruct the instructions and tool interactions behind a browser-agent failure. Pair them with test-run identifiers, application logs and browser traces to check whether the reported outcome matches the actual application state. A transcript alone does not establish that a test passed.
For an initial evaluation, run a controlled browser workflow with a known failure. Confirm that your export finds the correct product surface, retrieves every message page and preserves links to the original test artifacts. Allow for ingestion delay before treating a missing session as a capture defect.
Keep this beta separate from release-gate evidence until your team has checked coverage and completeness on representative runs. The announcement expands observability; it does not claim improved browser automation accuracy.