Anthropic Closes the Compliance Gap Around Claude Code and Cowork Sessions

Anthropic’s Compliance API, previously limited to Claude chat conversations, now covers Claude Cowork and Claude Code sessions in beta for Claude Enterprise customers. The new session endpoints return one consolidated, server-hosted transcript per session — prompts, responses, and tool activity, including web tools and MCP calls, alongside verified user information, organization IDs, and timestamps — instead of the fragmented logs security and eDiscovery teams previously had to stitch together across products. Coverage spans Cowork on desktop, web, and mobile, plus Claude Code in CLI and desktop environments; Claude Code on the web, direct Claude Platform API access, and sessions run through AWS Bedrock, Google Cloud Vertex AI, or Microsoft Foundry aren’t covered yet.

The timing tracks a real shift in how much unsupervised work these sessions now do. Auto mode became the default permission setting in Claude Code for Pro, Max, and Team plans on August 14 — routing tool calls through a safety classifier instead of requiring manual approval at every step. In testing, human reviewers caught only 13.6% of dangerous commands; auto mode blocked 89%. Teams running it shipped roughly 25% more pull requests, and one engineer described letting an agent run unattended overnight. That’s exactly the kind of session — longer, less supervised, doing more — that a compliance team can’t audit from a chat log alone.

It also follows Anthropic’s push into FinOps-style governance: the admin console added spend-threshold alerts, SCIM-group cost breakdowns, and Claude Code-specific usage and value tabs back in July. Between spend visibility and session-level audit trails, Anthropic is building the enterprise control layer around Claude Code roughly as fast as it’s expanding what Claude Code can do on its own — a pairing enterprise buyers evaluating agentic coding tools should expect to keep seeing.