openclaw/TELEMETRY.md

Telemetry

DenchClaw collects anonymous, non-identifiable telemetry data to help us understand how the product is used and where to focus improvements. Participation is optional and can be disabled at any time.

Telemetry is split into two independent layers:

1. Product telemetry — lightweight CLI and web-app usage events.
2. AI observability — LLM generation, tool call, and feedback tracking via PostHog's LLM Analytics (powered by an OpenClaw plugin).

Both layers share the same opt-out controls and privacy mode setting.

---

Product Telemetry

Event	When	Properties
cli_bootstrap_started	denchclaw bootstrap begins	version
cli_bootstrap_completed	Bootstrap finishes	duration_ms, workspace_created, gateway_reachable, web_reachable, version
chat_message_sent	User sends a chat message in the web UI	message_length, is_subagent
chat_stopped	User stops an active agent run	—
workspace_created	New workspace is created	has_seed
workspace_switched	User switches workspaces	—
workspace_deleted	Workspace is deleted	—
session_created	New web chat session is created	—
object_entry_created	CRM object entry is created	—
report_executed	A DuckDB report is executed	—
file_uploaded	A file is uploaded to the workspace	—
$pageview	User navigates within the web app	$current_url (path only, no query params with user data)

Every event includes baseline machine context: os (platform), arch, and node_version. A SHA-256 hash of the machine hostname + username (truncated to 16 hex chars) is used as the anonymous distinct ID — it cannot be reversed to identify you.

---

AI Observability

The posthog-analytics OpenClaw plugin captures LLM interactions as PostHog AI events. It is installed automatically during denchclaw bootstrap when a PostHog project key is available.

Event hierarchy

Session ($ai_session_id)
  └─ Trace ($ai_trace_id)           ← one per agent run
       ├─ Generation ($ai_generation) ← the LLM call
       ├─ Span ($ai_span)            ← each tool call
       ├─ Span ($ai_span)
       └─ ...

Events

Event	When	Key properties
$ai_generation	Agent run completes	$ai_model, $ai_provider, $ai_input_tokens, $ai_output_tokens, $ai_latency, $ai_total_cost_usd, $ai_tools, $ai_is_error
$ai_span	Each tool call completes	$ai_span_name (tool name), $ai_latency, $ai_is_error, $ai_parent_id
$ai_trace	Agent run completes	$ai_trace_id, $ai_session_id, $ai_latency, tool_count
survey sent	User clicks Like/Dislike in the web UI	$survey_response (1=like, 2=dislike), $ai_trace_id, message_id
dench_message_received	User sends a message (gateway-side)	channel, session_id, has_attachments
dench_session_start	Agent session begins	session_id, channel
dench_session_end	Agent session ends	session_id, channel
dench_turn_completed	Agent run completes	session_id, run_id, model

Privacy mode

By default, privacy mode is on. When privacy mode is enabled:

• $ai_input and $ai_output_choices are replaced with [REDACTED].
• Tool call parameters and results are not included in $ai_span events.
• Only metadata is captured: model name, token counts, latency, cost, tool names, and error flags.

When privacy mode is off, full message content and tool results are captured. API keys, tokens, and credential-like strings are always stripped regardless of privacy mode.

Toggle privacy mode:

denchclaw telemetry privacy off    # capture full content
denchclaw telemetry privacy on     # redact content (default)

PostHog evaluations

Once AI events are flowing, you can configure PostHog Evaluations in the dashboard to automatically score generations:

• LLM-as-a-judge — score outputs on relevance, helpfulness, hallucination, or custom criteria.
• Code-based (Hog) — deterministic checks like output length, keyword presence, or cost thresholds.

Evaluations run on sampled $ai_generation events and store pass/fail results with reasoning. No code changes are needed — evaluations are configured entirely in the PostHog dashboard.

User feedback (Like / Dislike)

The web UI shows thumbs-up and thumbs-down buttons on every completed assistant message. Clicking a button sends a survey sent event to PostHog linked to the conversation's $ai_trace_id. This feedback appears in the PostHog LLM Analytics trace timeline.

Feedback buttons only appear when the PostHog project key is configured. If PostHog is unreachable, feedback calls fail silently — the chat UI is never blocked.

---

What We Do NOT Collect

• File contents, names, or paths
• Message contents or prompts (when privacy mode is on — the default)
• API keys, tokens, or credentials (always stripped)
• Workspace names (never sent, not even hashed)
• IP addresses (PostHog is configured to discard them)
• Environment variable values
• Error stack traces or logs
• Any personally identifiable information (PII)

---

How to Opt Out

Any of these methods will disable telemetry entirely (both product telemetry and AI observability):

CLI command

denchclaw telemetry disable

Environment variable

export DENCHCLAW_TELEMETRY_DISABLED=1

DO_NOT_TRACK standard

export DO_NOT_TRACK=1

CI environments

Telemetry is automatically disabled when CI=true is set.

Check status

denchclaw telemetry status

---

Configuration

Privacy mode

denchclaw telemetry privacy on     # redact message content (default)
denchclaw telemetry privacy off    # send full message content

Privacy mode is stored in ~/.openclaw-dench/telemetry.json and is read by both the CLI/web telemetry layer and the OpenClaw analytics plugin.

PostHog analytics plugin

The plugin is configured via OpenClaw's plugin config:

openclaw --profile dench config set plugins.entries.posthog-analytics.enabled true
openclaw --profile dench config set plugins.entries.posthog-analytics.config.apiKey <key>

This is handled automatically by denchclaw bootstrap.

---

Debug Mode

Set DENCHCLAW_TELEMETRY_DEBUG=1 to print telemetry events to stderr instead of sending them. Useful for inspecting exactly what would be reported.

Re-enabling

denchclaw telemetry enable

How It Works

• CLI: The posthog-node SDK sends events from the Node.js process. Events are batched and flushed asynchronously — telemetry never blocks the CLI.
• Web app (server): API route handlers call trackServer() which uses the same posthog-node SDK on the server side.
• Web app (client): The posthog-js SDK captures pageview events in the browser. No cookies are set; session data is stored in memory only.
• OpenClaw plugin: The posthog-analytics plugin runs in-process with the OpenClaw Gateway. It hooks into agent lifecycle events (before_model_resolve, before_prompt_build, before_tool_call, after_tool_call, agent_end, message_received, session_start, session_end) and emits PostHog AI events via posthog-node.
• PostHog project token: The write-only project token (phc_...) is embedded in the built artifacts. It can only send events — it cannot read dashboards or analytics data.