A local coding session can feel like work between one developer and one laptop. Anthropic’s new enterprise export turns part of that work into a centrally retrievable compliance record—and the boundary is narrower and more sensitive than the feature name suggests.
What changed in the Claude Code transcripts API
Anthropic added local application sessions to its Compliance API on August 11, 2026. Enterprise organizations in the beta can list local sessions, retrieve one session, and fetch its messages through three read-only endpoints. The same Compliance Access Key used elsewhere in the API needs the read:compliance_user_data scope.
GET /v1/compliance/apps/lists local sessions.sessions/ local GET /v1/compliance/apps/returns session metadata.sessions/ local/ {session_id} GET /v1/compliance/apps/returns the transcript.sessions/ local/ {session_id}/ messages
This is not a remote-control interface. Anthropic documents it as a read-only export surface. It does, however, change the governance assumption around signed-in local work: eligible Claude Code transcripts can now enter the same audit pipeline as other enterprise records.
The export covers API-visible conversation, not the whole laptop
The scope includes Cowork Desktop and signed-in Claude Code sessions used in a terminal, the desktop app, or an IDE. Anthropic says the transcript reflects what was sent to the Claude API. Local commands, files, or device activity that never reached the API are not reconstructed by this endpoint.
That distinction prevents two opposite mistakes. Security teams should not treat the export as a complete endpoint-forensics record. Developers should not assume that “local” means invisible to the organization when the conversation was transmitted through an eligible enterprise account.
What the transcript contains—and what it omits
Anthropic says the exported messages can include user prompts, assistant text, tool calls, and the text portions of tool results. Thinking blocks, system prompts, tool definitions, and MCP configuration are omitted. Binary blocks are replaced rather than delivered as the original file.
The omission list is important, but it is not a sanitization promise. Anthropic explicitly warns that transcript content is not masked for URLs, credentials, or personal data. A copied access token, customer identifier, internal hostname, or sensitive tool result can therefore travel into the compliance archive.
The practical rule is simple: store these exports as sensitive security data, not ordinary analytics. Encrypt them, separate access from general reporting roles, and log every retrieval.
Retention can make a debugging shortcut a six-year record
Anthropic’s documentation says local sessions default to six years of retention. If an organization sets finite conversation-retention periods, the shortest applicable period governs. The API itself does not provide a deletion operation.
That should change onboarding language. A developer who pastes a production fragment to diagnose a one-minute problem may create a record that survives far longer than the immediate task. The policy should explain the retention period before use, identify prohibited data, and give developers an approved redaction or synthetic-data path.
Several Claude Code sessions remain outside this export
The beta does not cover every way to run Claude Code. Anthropic excludes sessions authenticated through a Claude Code Console API key, Amazon Bedrock, Google Cloud Vertex AI, or Microsoft Foundry. Claude Code on the web is also outside this local-session endpoint. Local sessions in HIPAA-ready organizations and those using zero-data-retention arrangements are excluded as documented.
A compliance dashboard that only counts exported sessions will therefore produce a false sense of completeness in a mixed environment. Inventory authentication routes and execution surfaces first. Then label which ones generate centrally retrievable Claude Code transcripts and which require a different evidence source.
A safer rollout uses purpose-limited access
I would not connect this feed directly to a broad data lake on day one. Start with a small, documented compliance purpose and a narrow reviewer group.
- Map coverage. Record which clients, login methods, organizations, and retention settings are included.
- Minimize collection. Export only the sessions required by a stated legal, security, or audit need.
- Separate duties. Do not give every engineering manager access to raw transcripts merely because the endpoint exists.
- Protect secrets. Scan the archive for credentials and personal data without treating scanning as a substitute for access control.
- Set an incident path. Define who can quarantine an exposed token, notify the owner, and document the response.
- Test retrieval. Verify that the session index, message export, and retention policy tell the same story before relying on the archive.
That approach matches the broader lesson from Claude Inference Hooks: putting an AI control in the critical path creates a new system that needs its own failure policy. It also complements our review of Claude Code cross-session messaging, where identity and permissions matter more than the convenience of the handoff.
My verdict: auditability is useful only with an honest boundary
The new export gives enterprise teams a useful record of eligible API-visible local work. It does not capture an entire machine, cover every Claude Code authentication route, or remove sensitive content before delivery.
Treat the API as a sensitive evidence pipeline. Tell developers what is retained, document the exclusions, restrict raw access, and test whether the archive can answer the compliance question you actually have. A transcript is valuable evidence; it is not a complete account of everything an agent did.
Read the primary documentation
- Read Anthropic’s Compliance API release note.
- Review the local session transcript documentation, including retention, exclusions, and sensitive-data warnings.
Source note: feature behavior and coverage in this article come from Anthropic’s documentation. Musthave.AI did not independently reproduce the Enterprise beta.
