Claude Code Runs on Your Laptop. The Transcript Lives 6 Years.

TL;DR

On August 11, Anthropic extended its Compliance API to cover Cowork and Claude Code sessions that run on a user's own machine. Three new endpoints let a Claude Enterprise compliance team list those sessions and pull the full transcript: prompts, model responses, tool calls, tool results, and the file contents Claude read along the way. Captured content is held for six years by default. The endpoints are read-only, so nobody, user or admin, can delete a session through them. This is beta, it is Claude Enterprise only, and it needs no new key, no new scope, no setting change, and no client update.

What actually shipped

The Compliance API is the thing security, legal, and eDiscovery teams already point at a Claude Enterprise tenant to pull chats, files, and projects. Until this week it stopped at claude.ai. Now it reaches the two surfaces where the interesting work happens: Cowork in Claude Desktop, and Claude Code in the terminal, in Claude Desktop, or in an IDE extension.

Three endpoints, documented in the content-data reference:

GET /v1/compliance/apps/sessions/local
GET /v1/compliance/apps/sessions/local/{session_id}
GET /v1/compliance/apps/sessions/local/{session_id}/messages

List, retrieve metadata, retrieve transcript. Sessions carry a product_surface of cowork or claude_code, so an admin can filter for exactly the terminal work. All three require the read:compliance_user_data scope on a Compliance Access Key, which is the same scope that already reads chats and files. Anthropic's docs are explicit that "no new key, scope, setting, or client update is required." If your organization already runs the Compliance API, the coverage arrived on its own.

Nothing is installed on your machine

Here is the part worth understanding properly, because it is the difference between a scandal and an inevitability. There is no agent on the laptop, no endpoint monitor, no new telemetry channel. Anthropic records the conversation server-side as its requests reach the Claude API. In the docs' words: "nothing is installed on the device, and nothing is collected beyond the requests the client already sends to the Claude API."

Think of it like the phone company. Nobody bugged the handset in your pocket. The handset was always fine. It is the line that remembers, and someone just published the number you dial to ask for the tape.

The practical consequence of that design: on-device activity that never reaches the API is not captured. If Claude never saw it, the transcript never had it. Grep a file yourself, and it stays yours. Ask Claude to read it, and the text lands in a tool result.

What lands in the transcript

Prompts come back as text blocks. Assistant responses come back as text output only. Tool activity comes back as tool_use inputs and tool_result text, truncated to 10,000 bytes per block by default, raisable to roughly 1 MiB per block on request via tool_use_input_max_bytes and tool_result_max_bytes. Text that Claude read through a tool appears in that transcript under the same cap. Artifacts show up inside tool-call inputs. Skill content appears when the client sends it as message content, undistinguished from any other user text.

What does not come back is almost as interesting. No thinking blocks. No tool definitions. No token usage, cost, or latency. No host or device metadata, so no terminal type and no workspace paths. Images, PDFs, and other binary content arrive as placeholder text blocks with truncated set to true, and raw file bytes are never returned. For usage and performance telemetry, Anthropic points admins at Claude Code's OpenTelemetry monitoring instead, which is a push model into infrastructure you run.

The six-year clock

Captured local session content is stored for six years from capture by default. If the organization has set a finite custom conversation retention period, that applies instead, shorter or longer, and where several are configured the shortest wins. One asymmetry to note: each captured message is stored under the period in force when it was captured, so lengthening the window later does not resurrect content that already expired.

Six years is a different order of magnitude from every other Claude Code retention number, and those numbers all live in the same docs set.

Meanwhile your own machine is the forgetful one. Claude Code caches session transcripts locally in plaintext under ~/.claude/projects/ for 30 days by default, tunable with cleanupPeriodDays. Your laptop moves on in a month. The compliance copy does not.

Read-only means read-only

The session endpoints support no delete. Chats, files, and projects can be hard-deleted through the Compliance API; sessions cannot. So the afternoon you spent losing an argument with a flaky test is now durable infrastructure, and there is no verb in the API for taking it back.

Where the coverage stops

The boundaries are real, and several of them are the ones a careful team would actually rely on:

• Auth path matters. Claude Code authenticated with a Claude Console API key is not captured. Neither is Claude Code running through Amazon Bedrock, Google Cloud, or Microsoft Foundry.
• Claude Code on the web is out. It runs in Anthropic-managed cloud environments, but the remote session endpoints return Cowork sessions only, so it falls in neither bucket.
• HIPAA readiness turns it off entirely. Organizations with HIPAA readiness enabled get no local session data at all.
• Zero data retention sessions are excluded, and the messages endpoint returns 404 for them.
• Customer-managed keys split the difference. Sessions list and retrieve as usual, but every message comes back with a content_unavailable provenance and an empty content array.

One quirk worth handling in code: the list is built from session activity metadata, so it can surface sessions that ran before capture began for your organization, as far back as retention allows, each one an empty shell marked not_captured. Do not read a listed session as a captured session. All /v1/compliance/* endpoints share 600 requests per minute per parent organization, and the local session endpoints get no extra budget on top, so a full historical export needs pacing.

Why a builder should care

Most of this was already true in substance. If you are signed in with an Enterprise account, your prompts have always crossed Anthropic's wire, and your employer has always been able to ask for records through legal process. What changed on Monday is the shape of the ask: it went from a request and a wait to a GET with a key the security team already holds, against a store with a six-year default.

The honest read is that this closes a governance gap that was blocking Claude Code from regulated environments, which is a good outcome for anyone who wants agentic coding approved at a bank. The equally honest read is that the terminal, the last place that felt like it was only yours, now has a documented retrieval endpoint. Both things are true, and the docs are refreshingly plain about it, which is more than most vendors manage.

If you care about the boundary, the levers are the auth path, ZDR, and the organization's custom conversation retention period. If you do not care, nothing changes, which is very much the point.

Key Takeaways

• Anthropic's Compliance API now returns transcripts of Cowork and Claude Code sessions that run on users' own machines, in beta for Claude Enterprise, announced August 11, 2026.
• Three read-only endpoints under /v1/compliance/apps/sessions/local list sessions and return full transcripts, using an existing Compliance Access Key with read:compliance_user_data.
• Captured content is retained six years by default, or the organization's custom conversation retention period; sessions cannot be deleted through the API.
• Capture is server-side at the Claude API, so nothing is installed on the device and on-device activity that never reaches the API is not captured.
• Transcripts carry prompts, responses, tool calls and results (10,000 bytes per block by default), and file text Claude read, but no thinking blocks, tool definitions, cost data, or workspace paths.
• Console API key auth, Bedrock, Google Cloud, Microsoft Foundry, Claude Code on the web, HIPAA readiness, and ZDR are all outside the capture boundary.

Sources: Anthropic, Compliance API coverage extends to Claude Cowork and Claude Code, Claude Platform Docs, Retrieve and delete chats, files, projects, and sessions, Claude Platform Docs, Compliance API FAQ, Claude Platform Docs, Compliance API, Claude Code Docs, Data usage, Claude Platform Docs, API and data retention