Skip to main content
Route complete audit telemetry from Claude for Excel, PowerPoint, Word, and Outlook to your own OpenTelemetry (OTEL) collector. This gives you control over retention and encryption and lets you feed the data into a SIEM or observability platform. This capability is available to Claude Enterprise organizations and to direct-provider deployments on Amazon Bedrock, Google Cloud Vertex AI, Azure AI Foundry, or an LLM gateway.

​
What the collector receives

When a custom collector is configured, the add-in sends trace data for every user turn. Each turn produces a tree of spans covering the prompt, each model call, each tool execution, file uploads, and context compaction. All span attributes are included. They carry user-generated content: prompt text, tool inputs and outputs, and document URLs. The allowlist filter that strips sensitive attributes on the Anthropic path does not run on this path. One normalization still applies: MCP connector tool names are recorded as the literal mcp_tool rather than the connector-specific name. Assistant response text is not included in span data. Your organization owns the collected data; treat the endpoint as containing prompt and document content when you scope access controls and retention. Metrics are not sent to custom collectors. The office_agent.* counter namespace routes to Anthropic only. Every counter increment is also recorded as a span event on the active span, so the same signals are recoverable from the trace stream. Telemetry is sent over OTLP/HTTP to {your_url}/v1/traces. gRPC is not supported because of Office WebView constraints.

​
Endpoint requirements

The add-in exports telemetry from each user’s browser: the taskpane at https://pivot.claude.ai posts directly to your collector, so every export is a cross-origin request. The endpoint must:
  • Answer the OPTIONS preflight with Access-Control-Allow-Origin covering https://pivot.claude.ai and Access-Control-Allow-Headers covering Content-Type plus any headers you configure, such as Authorization.
  • Return the same Access-Control-Allow-Origin header on the POST response.
Managed OTLP ingest endpoints such as Grafana Cloud are built for server-to-server export and generally do not answer browser CORS preflights. Point the add-in at an OpenTelemetry Collector you run, configure the cors block on its OTLP HTTP receiver, and have that collector forward to your backend. The collector makes the authenticated call to your backend, so no credential needs to appear in export headers that reach every signed-in user’s browser.

​
Set up the collector

Configuration differs by how your users sign in.

​
Claude Enterprise organizations

An organization administrator sets the collector endpoint in the Claude admin console under Organization settings, Office agents. Two settings are available:

​
Direct-provider deployments

Deployments that authenticate against Bedrock, Vertex AI, Foundry, or a gateway supply the same two keys through the customer-configuration channels described in Use Claude for M365 with third-party platforms. The claude-for-msft-365-install setup plugin writes these for you. To set them by hand, use any of the three channels below. Later channels override earlier ones: manifest parameters are read first, then Entra claims, then the bootstrap response. Manifest URL parameter. Append the keys to the taskpane URL in your custom manifest.
Entra ID directory extension. Register the keys as directory extension attributes and assign them per user through Microsoft Graph. The add-in reads them from the user’s ID token via Nested App Authentication. extn.otlp_endpoint maps to otlp_endpoint and extn.otlp_headers maps to otlp_headers. Bootstrap endpoint response. Include the keys in the JSON body your bootstrap endpoint returns.

​
Attribute size cap

Prompt, tool-input, and tool-output attributes are truncated in the add-in at 4,000 characters each by default and marked with a trailing …[truncated]. Set the otlp_attr_max_chars configuration key to a positive integer to change the cap. Values are clamped to between 256 and 32,000. Before raising the cap, confirm your collector and tracing backend accept attribute values of the configured size: many backends truncate or drop over-limit attributes at ingest, and a dropped span is lost from the audit trail entirely.

​
Deployment modes

The audit trail differs slightly depending on the sign-in path. Claude Enterprise (OAuth): full audit trail including user identity (user.email, user.account_uuid, organization.id), MCP server metadata, and file-upload spans. Direct provider (Bedrock, Vertex AI, Foundry, gateway): core audit trail with prompts, tool inputs and outputs, and document URLs. No Claude user identity, MCP metadata, or file-upload spans. Attribute activity to a user by correlating session.id against your identity provider or gateway logs.

​
Span reference

Each user turn produces up to five span types. agent.query is the root; agent.stream and agent.compaction are its children; agent.tool_execution is a child of agent.stream; file.upload arrives as a separate root span. The agent.query and agent.compaction spans carry agent.surface (sheet, doc, slide, or mail) and agent.vendor (m for Microsoft); the other three do not. To filter those by surface, join agent.stream and agent.tool_execution to their parent agent.query by trace ID, and correlate file.upload to a turn by session.id and timestamp, since it has its own trace ID and no surface attribute. Attributes marked content carry user-generated data. Attributes marked Claude sign-in only are populated only when users sign in with a Claude account.

​
Resource attributes

These are set on every span.

​
agent.query

Root span, one per user turn. SpanKind INTERNAL.

​
agent.stream

One span per model API call, child of agent.query. SpanKind CLIENT. The add-in requests prompt caching on every call. Cache token attributes are set from the provider’s response and omitted when the provider does not return them.

​
agent.tool_execution

One span per tool call, child of agent.stream. SpanKind INTERNAL. This is the primary record of what the model did to the document.

​
agent.compaction

One span per automatic conversation summarization when context nears the window limit, child of agent.query. SpanKind CLIENT. Also carries agent.surface, agent.vendor, session.id, office.platform, office.version, and user.email (Claude sign-in only).

​
file.upload

One span per uploaded file, emitted as its own root span rather than under agent.query. SpanKind CLIENT. Claude sign-in only. Also carries session.id and user.email. Correlate to the turn by session.id and timestamp.

​
Span events

Spans carry timestamped events for lifecycle transitions:
  • agent.query: exception, file_upload
  • agent.stream: first_token, stream_complete, stream_error
  • agent.tool_execution: tool_init, tool_run, tool_result, tool_error
  • agent.compaction: compaction_start, compaction_complete, compaction_error
  • file.upload: exception
Every internal product counter also records a span event with the same name on the active span. For example, office_agent.token.usage is emitted on each agent.stream span with token_usage.type (input, output, cacheRead, or cacheCreation), token_usage.model, and token_usage.token_count. Surface-specific events include office_agent.cell_edit_collision_total on Excel when a user is mid-edit while a tool writes, and the Word document-edit funnel (office_agent.doc_edit_received_total, doc_edit_parsed_total, doc_edit_applied_total, doc_proposed_edit_reviewed_total). PowerPoint and Outlook add no events beyond the common schema.

​
Reconstruct a user session

The span tree produces a complete, ordered transcript in both deployment modes. For Claude Enterprise deployments, filter agent.query spans by user.email or user.account_uuid and session.id, order them by timestamp, and read user.message and document.url for each turn. Then follow each agent.query span’s trace ID to its agent.tool_execution descendants, ordered by timestamp, to see what was attempted (tool.input), the result (tool.output), and how it was approved (tool.accept_decision). For direct-provider deployments, filter agent.query spans by session.id to isolate one session, use document.url to identify the file, and correlate the session against your Entra sign-in events, gateway access logs, or bootstrap endpoint logs to attribute it to a user. Per-turn reconstruction then follows the same trace-ID walk.