Monitoring is available for Team and Enterprise plans. OTel monitoring requires Claude desktop app version 1.1.4173 or later.
Setup
Configure monitoring from Organization settings > Data and privacy:- Go to Organization settings > Data and privacy
- Under Monitoring, find the Cowork and Cowork in the cloud item
-
Configure the following fields:
- Save your settings
- Start a new Cowork session. Cowork loads settings when a session starts, so sessions that are already running keep the old configuration.
By default, Cowork runs the OTel exporter on the user’s device and sends events over that device’s network connection. Make sure your collector is reachable from users’ devices.You don’t need to add your collector to the Additional allowed domains field in Organization settings > Capabilities. If sessions run in the Cowork VM and your organization restricts network egress, Cowork adds your collector’s hostname to the session’s allowlist for you.
Events
Cowork exports the following events to your OTel collector. The collector settings in Setup apply to every Cowork session that runs on a user’s computer in Claude Desktop, including Cowork tasks that Dispatch starts. They don’t apply to Code tab sessions, including Code sessions that Dispatch starts. Whether events include user prompt text, model response text, and tool inputs depends on theotlpContentCapture managed configuration key on each user’s device. When the key isn’t set, current versions of Claude Desktop send metadata only, and some older versions on first-party deployments send content.
Content capture
Events can carry user prompt text, model response text, and tool inputs in addition to metadata. You set which of these the events carry with theotlpContentCapture managed configuration key on each user’s device.
To set the key, deploy it with your device management tool to the managed configuration location for each operating system. To export metadata only, set otlpContentCapture to an empty list, written as the two characters []. Claude Desktop reads an empty string as if the key weren’t set. The key takes effect when Claude Desktop restarts.
The table shows what events carry on current versions of Claude Desktop, for a Cowork session that runs on the user’s computer.
On first-party deployments, some older versions of Claude Desktop send user prompt text, model response text, and the
toolDetails content when the key isn’t set. To make sure a device sends metadata only, set the key to [], which works on every version that supports the key.
Metadata includes workspace.host_paths and, on first-party deployments, user.email. The key doesn’t control either one. See Security and privacy.
Event correlation
When a user submits a prompt, Cowork may make multiple API calls and run several tools. Theprompt.id attribute links all events back to the single prompt that triggered them.
To trace all activity triggered by a single prompt, filter your events by a specific
prompt.id value.
To order a session’s events, sort by event.timestamp and use event.sequence to order events that share a timestamp. A session can stop running between prompts, for example when a task sits idle or Claude Desktop restarts. When the session resumes, it keeps its session.id and event.sequence starts over. A later event in the session can therefore carry a lower value than an earlier one, or repeat one.
On third-party deployments, you can additionally enable OpenTelemetry trace export with the otlpTracesEnabled setting (beta). When it is enabled, events emitted while a prompt is processed also carry trace_id and span_id, linking them to the session’s trace spans for end-to-end correlation in your observability backend.
Standard attributes
All events include these attributes:The account attributes (
organization.id, user.account_uuid, user.account_id, and user.email) are populated from the user’s Anthropic account, so they appear on first-party deployments only. On third-party deployments there is no Anthropic account and these attributes are absent; instead, the export carries the signed-in user’s identity as the enduser.id resource attribute, described under User attribution. The process.owner resource attribute (the operating-system login name) is standard OpenTelemetry process metadata and is present on all deployments.User prompt event
Logged when a user submits a prompt. Event name:user_prompt
Attributes:
All standard attributes, plus:
Model response event
Logged when the model completes a response that includes text output. Requires Claude desktop app version 1.17377 or later. Event name:assistant_response
Attributes:
All standard attributes, plus:
By default, model responses are captured when
otlpContentCapture includes assistantResponses, and also whenever user prompts are captured.
Tool result event
Logged when a tool completes execution. Event name:tool_result
Attributes:
All standard attributes, plus:
API request event
Logged for each API request to Claude. Event name:api_request
Attributes:
All standard attributes, plus:
Claude Desktop 3.1623.0 and later send
cache_creation_5m_tokens and cache_creation_1h_tokens. Both are 0 when the API response doesn’t report the split, even when cache_creation_tokens is above zero. If your deployment uses an LLM gateway, confirm that the gateway returns the cache_creation object in each response’s usage.
API error event
Logged when an API request to Claude fails. Event name:api_error
Attributes:
All standard attributes, plus:
Tool decision event
Logged when a tool permission decision is made. Event name:tool_decision
Attributes:
All standard attributes, plus:
Cowork shows Auto mode as Automatically approve. An event has
classifier_verdict when the Auto mode classifier decides the call, or when Cowork denies the call because the classifier gives no verdict. An event with classifier_verdict reports source as "config". The attribute takes one of these values:
"allow": the classifier approves the call"block": the classifier blocks the call. A check that fails partway through can also end as"block"."no_verdict": the classifier’s check runs but gives no usable verdict, so Cowork denies the call"unavailable": the check can’t run or can’t finish, for example because the request to the classifier fails, so Cowork denies the call
source and classifier_verdict value covers, and for the Auto mode calls that have no classifier_verdict, see Tool decision event in the Claude Code documentation. Cowork tasks follow what that section says for Agent SDK sessions.
Claude Desktop 3.1623.0 and later send permission_mode, classifier_verdict, and classifier_category. Claude Desktop 4.16336.0 and later send request_id.
Event analysis
The exported events support a range of analyses: Tool usage patterns: Analyze tool result events to identify most frequently used tools, success rates, average execution times, and error patterns. Cost monitoring: Trackcost_usd from API request events to understand usage trends across users and teams. Group by user.account_uuid or organization.id for per-user or per-team breakdowns.
Performance monitoring: Track API request durations and tool execution times to identify performance bottlenecks.
Cost values from events are approximations. For official billing data, refer to your billing dashboard.
Backend considerations
Your choice of logs backend determines the types of analyses you can perform:- Log aggregation systems (e.g., Elasticsearch, Loki): Full-text search and log analysis
- Columnar stores (e.g., ClickHouse): Structured event analysis and complex queries
- Observability platforms (e.g., Honeycomb, Datadog): Advanced querying, visualization, and alerting
Service information
All events are exported with the following resource attributes:Security and privacy
- Events are only exported when an admin configures the OTLP endpoint
- The
otlpContentCapturekey on each device sets which content events carry. Each category you list in the key adds its content. The full category list has two more than these three:userPrompts: user prompt textassistantResponses: model response texttoolDetails: the arguments a tool was called with, intool_inputandtool_parameters, such as shell commands, file paths, URLs, and search patterns. It also covers the text of a failed tool’s error message, inerroron the tool result event, and the classifier’s rule name inclassifier_categoryon the tool decision event
- On Claude Desktop version 1.17377 or later, events that carry user prompt text also carry model response text by default, even when the key doesn’t list
assistantResponses - When
otlpContentCaptureisn’t set, current versions of Claude Desktop send metadata only. To make sure a device sends metadata only, set the key to[] - Metadata includes
workspace.host_paths, the paths of the folders a user connects to a task.otlpContentCapturedoesn’t control this attribute, so events carry the paths even when the key is[]. If folder names can be sensitive, configure your telemetry backend to filter or redact it - On first-party deployments,
user.emailis always included in event attributes, so configure your telemetry backend to filter or redact it if this is a concern - On third-party deployments,
user.emailis absent; the export identifies users with theenduser.idresource attribute, controlled by theendUserAttributionsetting