English
All articles
Claude CodeFixesSetup & config

Claude Code telemetry: check Datadog traffic, privacy and permissions

Claude Code telemetry self-check: verify Datadog destinations, test DISABLE_TELEMETRY, and reduce permissions. A public report is not a confirmed ban mechanism.

6 min read
Pastel clay laptop, connected nodes and shield on cream with leaf shadows
Also available in中文Español

Check the destinations Claude Code reaches, test the telemetry opt-out, and trim unnecessary permissions. This Claude Code telemetry checklist separates a researcher’s report from Anthropic’s documentation. It does not establish why any account was banned.

Checked on November 6, 2026, Beijing time, UTC+8. The official changelog currently starts at 2.1.318. This article uses MinLiBuilds’ public report and official docs. We did not reproduce his 24-hour capture or inspect the reported request bodies.

Claude Code telemetry self-check before you change anything

  • Record your Claude Code version, account type, model provider, launch method, and enabled MCP servers and plugins. A terminal session and an IDE session can have different network activity.
  • Use a disposable workspace without secrets. Record process IDs, child processes, destination hostnames, times, and the action that caused each connection.
  • Compare the same short task before and after an opt-out. Restart Claude Code with the setting applied. Keep authentication and model access working.
  • Review /permissions, additional directories, hooks, MCP credentials, and browser access. Remove access the task does not need.
  • Keep raw logs private. Redact tokens, cookies, account identifiers, paths, and project names before sharing evidence.

What MinLiBuilds observed, and what remains a claim

MinLiBuilds posted on November 5 at 21:41 Beijing time, UTC+8. He says a local Qwen-built process-level monitor ran for 24 hours and saw four destination hosts. His attached image lists the following hosts. The descriptions here come from Anthropic’s network documentation.

Host in his imageDocumented use
api.anthropic.comModel API requests, WebFetch domain safety checks, feature flags, and telemetry event logging.
http-intake.logs.us5.datadoghq.comOptional operational telemetry when the CLI connects directly to the Anthropic API.
bridge.claudeusercontent.comThe WebSocket bridge for the Claude in Chrome extension.
downloads.claude.aiNative installation, updates, version checks, and plugin executable downloads.

He also reports device and session identifiers, account and organization UUIDs, model version, system environment, subscription type, Git commit SHA, and workspace state in requests. We have not independently verified those payloads. His post does not establish that every listed field went to Datadog, or that all installations contact only four hosts. The official network list is longer and depends on enabled features.

Datadog Claude searches: check the destination and the data separately

The Datadog host is documented. A connection to it alone does not tell you the request body. Anthropic’s data usage page says operational metrics cover latency, reliability, and usage patterns, and never include code, prompts, or file paths. It separately describes internal error reports, with redaction of known secret and personal-data patterns.

Do not apply the operational-metrics statement to every data channel. Model requests carry conversation data. Optional feedback can include transcripts and code. Organization-configured OpenTelemetry export is another channel, enabled with CLAUDE_CODE_ENABLE_TELEMETRY=1. Its documented attributes include session and account identifiers. Extra logging settings can include prompts, tool input, and commit details. That schema is not proof that the same fields went to Datadog in this capture.

Check which addresses your own process reaches

On macOS or Linux, if lsof is installed, this read-only check lists a chosen process’s open network sockets. Use your operating system’s process viewer to find its PID. On Windows, use the process-specific network view in Resource Monitor.

claude --version
# Replace 12345 with the Claude Code process ID.
lsof -nP -a -p 12345 -i

A socket snapshot can miss brief requests, and -n leaves addresses numeric. Use an approved process-aware network monitor or your organization’s proxy logs for a longer record with destination hostnames. Include child processes and, for an IDE session, the extension host. Encrypted connections reveal destinations and timing, not plaintext request fields. Do not install an unknown interception certificate to investigate a social post.

Repeat a small task in the same workspace and include shutdown in the capture. Compare startup, an idle interval, one model request, and exit. Label any MCP or browser action separately. A changed host count is an observation about that test, not a universal endpoint list. Issue #103435 reports different proxy behavior for the CLI and the IDE extension host on 2.1.301. It is a user report, not an Anthropic policy statement or proof that 2.1.318 behaves the same way.

Disable optional telemetry, then retest

For a terminal-only test on macOS or Linux, start a new process with both documented opt-outs. This changes the environment for that command only.

DISABLE_TELEMETRY=1 DISABLE_ERROR_REPORTING=1 claude

DISABLE_TELEMETRY=1 opts out of operational metrics. DISABLE_ERROR_REPORTING=1 opts out of error reports. For a broader test, the docs offer this alternative.

CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 claude

The broader switch disables non-essential traffic. It does not make Claude Code offline. The docs say the WebFetch domain safety check and official marketplace auto-install have separate controls. Disabling telemetry also disables feature-flag evaluation and can make Remote Control unavailable. Do not turn off safety checks merely to reduce a hostname count.

For an IDE or a persistent setup, apply the variables through the appropriate settings.json env block or approved launcher, and restart the processes that launch sessions. Do not overwrite an existing settings file. Check with your administrator if managed settings apply. Review any organization-configured OTEL_* exporters separately, especially content-logging flags and collector destinations. Never paste exporter headers into a bug report.

Retest the same task and compare the optional logging hosts. A blocked connection may still be an attempted connection. A missing connection in one short run does not prove that no data can ever leave. Model traffic, optional feedback, account training preferences, and retention policies need separate checks.

Reduce permissions before trusting a workspace

  • Open /permissions and remove broad allowances you no longer need, such as Bash(*) or an entire MCP server. Review the source of each rule.
  • Limit working directories. Review both read and write access. On supported versions, permissions.blockReadsOutsideWorkingDirectories and permissions.blockWritesOutsideWorkingDirectories restrict file tools. They do not confine arbitrary shell programs or MCP tools.
  • Use an OS sandbox, container, or VM for unfamiliar projects. Avoid running without permission prompts on your everyday machine. File rules alone are not an OS security boundary.
  • Review each MCP package’s publisher, version, startup command, and update source. Give it a scoped credential and avoid exposing production secrets. Inspect hooks and apiKeyHelper commands before trusting a repository.
  • Grant browser, Git push, deployment, and payment access only when the task needs them. Turning off telemetry does not revoke any of these permissions.

These checks follow the official permissions documentation. MinLiBuilds also connects his concern to GCSA’s Bitkub MCP warning and a Binance incident writeup. Those are separate security leads. We have not verified those incident details for this article, and they do not establish a Claude Code ban mechanism.

Does telemetry explain a Claude Code account ban?

The public report offers a hypothesis: metadata and request origin may be useful for risk assessment without reading chat content. We did not find an official response to this specific report in the sources reviewed. We have no verified evidence that Anthropic uses these reported fields to ban accounts, or that disabling telemetry prevents a ban. A 401, an expired OAuth token, and an account suspension also need different checks.

For the next step, use the Claude Code error hub, the 401 and OAuth guide, or the apiKeyHelper guide. Keep the exact error text and sanitized evidence. Do not turn a network observation into a ban verdict.

Sources and evidence limits