Anonymous telemetry

Patter sends a small amount of anonymous usage telemetry so the maintainers can see how the SDK is used — which engines, providers, models, carriers, platforms, and versions are popular, plus coarse per-call latency, duration, and cost — and prioritise accordingly. It is on by default and designed to be invisible to your application and impossible to tie back to a person or a call.

Telemetry never collects PII or call content. No phone numbers, transcripts, audio, prompts, tool arguments, API keys, customer identifiers, file paths, hostnames, or IP addresses are ever sent.

What is collected

Each event is a small, flat JSON object with only coarse, low-cardinality fields: Common fields on every event:

Field	Example	Notes
`event`	`sdk_initialized`	One of a fixed, small set of event names
`sdk` / `sdk_version`	`python` / `0.6.3`	Language and SDK version
`os` / `arch`	`linux` / `arm64`	OS family and CPU arch only
`runtime` / `runtime_version`	`node` / `20.11`	Bucketed to `major.minor`
`run_id`	random	A per-process random id, regenerated every run
`install_id`	random	A random, anonymous install id persisted locally, to count active installs (no PII, no hardware)
`ci`	`false`	Whether a CI/test environment was detected

Per-event dimensions (all are a closed enum or a coarse bucket — an off-list value is replaced with other before sending):

Event	Dimensions
`sdk_initialized`	`carrier` (twilio/telnyx/plivo/none), `tunnel` (static/configured/none), and anonymous deploy-shape (presence-only env/file probes — never the value of an env var): `invoked_by_agent` (claude/cursor/copilot/…/none), `container` (bool), `serverless` (lambda/cloud_run/vercel/azure_functions/none), `cloud` (aws/gcp/azure/fly/none), `package_manager`, `days_since_install_bucket`, and `previous_sdk_version` (for the upgrade funnel)
`first_run`	Sent once per install (the run that creates the local state), to mark activation. Carries the same anonymous deploy-shape dimensions as `sdk_initialized`. Never sent if you have opted out (opting out never touches the filesystem)
`cli_command`	`cli_command` — which `getpatter` CLI command was invoked (`dashboard` / `eval` / `hermes` / `openclaw` / `telemetry` / `other`). Only the command name — never arguments or flag values. The `getpatter telemetry` control command never emits telemetry itself
`feature_used`	`engine` (realtime/convai/pipeline), `provider` (vendor family), and the models in use: for pipeline the composed stack — `stt_provider` / `tts_provider` / `llm_provider` (vendor) plus `stt_model` / `tts_model` / `llm_model` (e.g. `deepgram-nova-3`, `anthropic-claude-opus-4-8`) — and for realtime the Realtime model as `llm_model` (e.g. `openai-gpt-realtime-2`). A fine-tuned, self-hosted, or custom model collapses to `{vendor}-other` — a custom model name is never sent
`agent_configured`	`builtin_tool_count` (0–3), `custom_tool_count_bucket` (a range like `2_3` — never the tool names), `integration` (openclaw/mcp/hermes/other/none), `integration_kind` (consult/mcp/none), `mcp_server_count_bucket`, and feature-adoption flags: `noise_reduction` (near_field/far_field/none), `turn_detection` (default/custom/none), `preambles_used` (bool), `per_tool_timeouts_set` (bool), `llm_fallback_configured` (bool)
`call_started`	Sent when a call connects (media stream begins), to pair with `call_completed` for a connect→complete funnel: `engine`, `provider`, `carrier`, `direction` (inbound/outbound), and `call_uid` — a random correlation id generated by the SDK so the two lifecycle events of the same call can be paired. It is never the carrier call SID and cannot be linked to phone numbers or call content. No metrics exist yet at connect, so no latency/cost/duration
`call_completed`	`outcome` (completed / error / no_answer / busy / failed), `direction` (inbound/outbound), `error_code` (a closed code like `rate_limit` / `timeout` / `connection` — never the error message), `engine`, `provider`, `carrier`, `latency_ms` (whole milliseconds), `duration_seconds` (whole seconds), `cost_usd` (total call cost in USD), `turn_count_bucket`, `call_uid` (same random correlation id as the call’s `call_started`) — sent once per call. No call content. The only per-call identifier is the SDK-generated random `call_uid` — never the carrier call SID

The run_id is a fresh random value on every process start (it groups one run’s events). The install_id is a random UUID generated once and stored in a small local file (~/.getpatter/install-id, or $XDG_STATE_HOME/getpatter/install-id when XDG_STATE_HOME is set) so the maintainers can count unique active installs over time — it is a random number, not tied to you or any PII, and is never derived from hardware (MAC address, hostname, serial numbers). Opting out never creates or reads this file. The collector drops the source IP on receipt. Custom tool names, custom integration endpoints, custom/fine-tuned model names, and any value not on a closed enum are never sent — only counts and coarse categories.

How to disable it

Telemetry is opt-out. Any one of the following turns it off:

Python

# In code:
phone = Patter(carrier=Twilio(...), phone_number="+15551234567", telemetry=False)

TypeScript

// In code:
const phone = new Patter({ carrier: new Twilio(), phoneNumber: "+15551234567", telemetry: false });

Or from the command line — a persisted, machine-level opt-out (parity with next telemetry disable):

getpatter telemetry disable   # persisted opt-out — no usage data is ever sent
getpatter telemetry enable    # re-enable (opt-out model, on by default)
getpatter telemetry status    # show the current state, endpoint, and how to audit

Or via environment variables (these also work without touching code):

Variable	Effect
`PATTER_TELEMETRY_DISABLED=1`	Disable Patter telemetry
`DO_NOT_TRACK=1`	Disable telemetry in Patter and every tool that honors the DO_NOT_TRACK convention

Telemetry is also automatically disabled in CI and test runners (for example when CI, GITHUB_ACTIONS, PYTEST_CURRENT_TEST, or VITEST is set).

Inspect exactly what would be sent

Set PATTER_TELEMETRY_DEBUG=1 to print every event to stderr without sending it. This lets you audit precisely what would leave your machine:

PATTER_TELEMETRY_DEBUG=1 python app.py
# [patter telemetry] {"event":"sdk_initialized","sdk":"python","sdk_version":"0.6.3",...}

Send telemetry to your own collector

Point Patter at any HTTP endpoint that accepts a JSON array of events:

export PATTER_TELEMETRY_ENDPOINT="https://telemetry.example.com/v1/ingest"

Precedence

When more than one signal is present, the first match wins:

DO_NOT_TRACK set → off
PATTER_TELEMETRY_DISABLED set → off
persisted opt-out marker present (getpatter telemetry disable) → off
telemetry=False in code → off
CI / test environment detected → off
otherwise → on

Get Started

Setting up Patter

Observability

Integrations

Development

What is collected

How to disable it

Inspect exactly what would be sent

Send telemetry to your own collector

Precedence

​What is collected

​How to disable it

​Inspect exactly what would be sent

​Send telemetry to your own collector

​Precedence

What is collected

How to disable it

Inspect exactly what would be sent

Send telemetry to your own collector

Precedence