Anthropic Compliance API

Claude Enterprise organisations can read transcripts of their users’ Claude sessions through Anthropic’s Compliance API. The connector pulls those sessions on a schedule and writes one decision record per session snapshot, so every Claude session in the company sits in the same signed, subject-indexed stream as every other AI system.

Requirements

• A Claude Enterprise organisation. Standalone Console organisations cannot use the session endpoints.
• A Compliance Access Key with the read:compliance_user_data scope, created in claude.ai by an organisation owner. A regular Claude API key (sk-ant-api03-…) does not work: the Compliance API answers it with a bare 404.

Set-up

1. Put the key where the Stratl API can read it, for example ANTHROPIC_COMPLIANCE_ACCESS_KEY.
2. Sources → Add connector → Anthropic Compliance API, secret reference env:ANTHROPIC_COMPLIANCE_ACCESS_KEY. Choose the session families: local (sessions on users’ machines: Claude Code, Cowork, Claude Science, Microsoft 365, Chrome) and remote (Cowork sessions started on claude.ai that run in Anthropic’s cloud).
3. Test, then Save. The first run looks back seven days; afterwards local sessions are polled by last activity with a ten-minute overlap, as Anthropic recommends, and remote sessions by creation time.

Some organisations receive Local sessions are not available while remote sessions work; the test reports that as a working key with local sessions unavailable.

What a session becomes

In the transcript	In the record
A user message	input, fingerprinted; content: not_captured with the reason when Anthropic withholds it
An assistant message	model_call with the serving model and a fingerprint of the text; the provenance marker (client_asserted, synthetic_marker) is kept
A tool_use block	tool_call with a fingerprint of the input, matched to its tool_result (fingerprint, error flag); truncation flags are kept
The session	One outcome (assistant_session); the session id is the subject; the product surface (Claude Code, Cowork, …) is the agent; the user is a human actor with identity_provider: anthropic

The decision id is derived from the session id and its last update, so a session that gains messages produces a new snapshot record and an unchanged session produces none. Prompts, responses and tool payloads are fingerprinted at the edge; nothing readable enters the record. Tool inputs and results are capped at 10 000 bytes per block by default (tool_bytes; -1 asks the server for its maximum).

Prove it against your account

ANTHROPIC_COMPLIANCE_ACCESS_KEY=sk-ant-api01-… make prove-compliance

Pulls, pulls again from the cursor, checks every record, writes nothing. Exit status zero is the proof. As with Langfuse, the connector is replay-tested against recorded responses matching Anthropic’s current API reference and has not yet been run by us against a production organisation.