Anthropic's compliance API now returns transcripts of Claude Code run on staff machines
The endpoints are in beta for Claude Enterprise organisations and need no new key, scope or client update. What is captured is what the client already sent to the API — nothing is installed on the laptop.
Anthropic's API changelog on 11 August added three endpoints that let a Claude Enterprise organisation list the Cowork and Claude Code sessions its people run on their own machines and pull the transcript of any one of them. They are in beta, and they work with the Compliance Access Key an organisation already holds: no new key, scope, setting or client update is required.
A transcript is the conversation as the API received it — the user's prompts, Claude's replies, every tool call with its input, and the text of every tool result. Project files such as CLAUDE.md appear as ordinary user content. Anthropic's own documentation is blunt about what that means: nothing masks URLs, credentials or personal data in that content, so transcripts should be treated as sensitive.
| Content of a local session | Returned |
|---|---|
| Everything the user typed | yes |
| Everything Claude replied | yes |
| Every tool call, with its input | yes |
| The text of every tool result | yes |
| CLAUDE.md, as ordinary user content | yes |
| Thinking blocks | no |
| The system prompt — a marker stands in for it | no |
| Tool definitions and MCP configuration | no |
| Images, PDFs and other binary blocks | no |
| Anything that never reached the API | no |
Six situations produce no local session data at all:
- Claude Code signed in with a Console API key
- Claude Code through Amazon Bedrock, Google Vertex or Microsoft Foundry
- Claude Code on the web
- Organisations with HIPAA readiness enabled
- Sessions under zero data retention
- Transcript text, in organisations using customer-managed keys
The limit is worth stating precisely, because it is narrower than an employer being able to see the laptop. Anthropic records each conversation server-side as its requests reach the Claude API. Nothing is installed on the device, nothing is collected beyond what the client already sends, and activity that never reaches the API — a local file the session never read out — is not in the transcript. Thinking blocks are never included, the system prompt is replaced by a marker, and images and PDFs come back as placeholders.
Several things are not captured at all. Claude Code signed in with a Console API key is out, as is Claude Code running through Amazon Bedrock, Google Cloud or Microsoft Foundry, and Claude Code on the web. Organisations with HIPAA readiness enabled capture no local session data, sessions under zero data retention are excluded, and organisations using customer-managed encryption keys can list sessions but not read their contents.
Retention is six years by default, or the organisation's own conversation retention period where it has set a finite one. The endpoints are read-only: a session cannot be deleted through this API, and capture is tied to the Compliance API being enabled for the organisation and to the user being signed in with their Enterprise account.
Under the heading of compliance this is unremarkable — the endpoints exist for eDiscovery and data loss prevention, and a company that already retains its employees' chat logs has been reading conversations for years. What is new is the surface. The sessions in question run on a developer's own machine, in a terminal, and the transcript records what an agent was asked to do and what it did.
The same changelog carried a second decision: the introductory price for Claude Sonnet 5, $2 and $10 per million tokens, is now the standard price, and the increase to $3 and $15 that had been scheduled for 1 September will not happen.
Sources
This article was written from these pages. Read them.
- primaryClaude API release notes — 11 August 2026docs.claude.com
- primaryRetrieve session transcripts — Compliance APIplatform.claude.com
Written from verified primary sources by Epoch's editorial pipeline and checked by a human before publication.