Claude Code Data Collection — Status Report (July 2026)

1What's Collected — and Where It's Sent

The whole picture, at a glance

Claude Code sends two separate kinds of data to two separate places. Operational telemetry (metadata about your work) goes to a third-party analytics service, Datadog. Your actual content — prompts, code, conversations — goes to Anthropic's servers and, on personal plans, is used to train the model. Here is every key item and exactly where it lands.

Datadog third-party analytics Sentry error reporting Anthropic vendor servers / training Local stays on your machine

Data being collected	What it exposes	Sent to
Account email & org/user IDs	Ties the whole stream to you — not anonymous	Datadog
Working directory paths	The shape of your filesystem & project names	Datadog
Repository & branch names	Your git repo identity and branch names	Datadog
Tool names, timing, exit codes	The rhythm of your work: what ran, how long, pass/fail	Datadog
Model, token counts, message length	How long your prompts were — a number, not the words	Datadog
Crash & error reports	Stack traces; may incidentally include a path fragment	Sentry
Your full prompts & conversations	Every word you type and receive, verbatim	Anthropic
Your code & coding sessions	The actual source you paste, generate, or edit	Anthropic
Files & context you share	Anything you attach or reference in a session	Anthropic
→ Used to train future models (personal plans, default ON)	Your content becomes training data for the next model	Anthropic
Tool output / results	Not sent to telemetry — measured only (exit code, token count)	Local

⚠ Personal plans (Free, Pro, Max): the content rows are ON by default Since a policy change in late 2025, personal plans have the data-training setting turned ON by default — an opt-out, not opt-in. If you never touched it, assume the Anthropic rows above are live: your conversations and code are being retained and used to train the model, with retention stretched from ~30 days to up to 5 years. The Datadog rows are on by default on all plans until you disable telemetry.

✓ Two separate switches The Datadog rows are killed by telemetry environment variables. The Anthropic content/training rows are killed by a separate account privacy setting (or by using a protected plan). Turning one off does nothing to the other — both sections below.

2Training Exposure by Plan

Who is protected, who is not

Whether the Anthropic content rows above are used to train the model depends entirely on your plan. Personal plans default ON; commercial plans are protected.

How it breaks down by plan

Plan	Is your content used for training?
Free	On by default Your chats and coding sessions are used to train the model unless you opt out. Retention extends to up to 5 years.
Pro	On by default Same as Free — content is used for training unless you switch the setting off. The burden is on you. Retention up to 5 years.
Max	On by default Same as Free and Pro — a paid personal plan does not exempt you. Training is on unless you opt out; retention up to 5 years.
Team	Protected Not used for training by default under commercial terms — unless your organization opts in to a data-sharing program.
Enterprise	Protected Not used for training by default. Zero-Data-Retention terms available for qualified accounts. The recommended tier for sensitive work.
API / Business (incl. Bedrock, Vertex)	Protected Not used for training by default under commercial terms — unless you explicitly opt in to a development-partner program.

The takeaway Privacy of your content is an account & setting question, not a telemetry question. Use a commercial / business tier for real work, and — because the toggle now defaults ON for personal plans — actively verify the data-training setting is off rather than assuming it. A personal plan used for work with the default left in place is the single highest-exposure configuration.

Watch the side channels, too

Beyond the two main streams, individual features can send content on demand — for example, a “send feedback” or “share this session” command that uploads your conversation, or an optional post-session survey that offers to upload a transcript. These are opt-in per use, but staff should know that clicking “yes” ships content regardless of your other settings.

3Configuring for Maximum Privacy

The switches — both streams

Maximum privacy means addressing both streams plus the side channels. Here is the checklist we apply for our clients.

A. Shut off operational telemetry

Most tools honor environment variables that disable the metadata stream at the process level. Set them where your tool runs so every session inherits them. Common industry-standard switches:

Environment variables (representative)

DISABLE_TELEMETRY=1 DISABLE_ERROR_REPORTING=1 DISABLE_BUG_COMMAND=1 DISABLE_NONESSENTIAL_TRAFFIC=1 # umbrella: covers the above + updater + experiments + surveys

Trade-off to know: an “umbrella” flag that blocks all non-essential traffic often also disables the auto-updater. On a controlled server that is arguably a feature — no silent version changes — but it means updates become a deliberate, manual step.

B. Lock down the content / training setting

Use a commercial / business / enterprise account for any real work — not a personal consumer login.
Open the account privacy dashboard and confirm the data-training toggle is OFF (steps below).
For the highest bar, ask your vendor about Zero Data Retention (ZDR) terms for qualified accounts.

🔑 Turn it off in Claude's settings (personal accounts)

Open your Privacy / Data controls and switch these two off:

Toggle Help improve our AI models OFF — stops your chats and code from being used to train the model (and shortens retention).
Toggle Location metadata OFF — stops coarse location data from being attached to your usage.

Open Claude Data & Privacy Controls →

C. Handle the side channels

Disable in-app feedback / session-share commands if the tool allows it, or train staff never to use them on sensitive projects.
Turn off optional post-session “share transcript” surveys.
Remember: a separate desktop app from the same vendor usually has its own telemetry toggle in Settings → Privacy — the command-line switches do not touch it.

Deployment order that actually sticks

Set the telemetry environment variables at the service / system level so every subprocess inherits them — a shell export alone often will not reach background workers.
Restart the service cleanly so the new environment takes effect.
Confirm the account data-training toggle is off, on a commercial tier.
Verify: check for outbound connections to the telemetry host after a real run — silence confirms success.

Bottom line

Your secrets, code, and prompts are safe from telemetry — but they are not automatically safe from model-training. Those are two different doors. Close both.

Kukui IT configures these controls for our managed clients as part of a standard secure-development posture. If you would like us to audit and lock down the AI tooling in your environment, reach out to your account team.

Sources

Primary references · verified July 2026

Facts in this report are drawn from Anthropic's own documentation and public reporting on the late-2025 consumer-terms change. Settings, defaults, and terms can change — always confirm against the current source.

Anthropic — Claude Code: Data usage — telemetry vs. training, retention, and the environment-variable opt-outs.
Anthropic — Updates to our Consumer Terms and Privacy Policy — the late-2025 change; consumer (Free/Pro/Max) plans included, commercial plans excluded.
Anthropic Privacy Center — Is my data used for model training? — the per-plan training and retention breakdown.
Anthropic — Data privacy controls — where consumer users turn the data-training setting off.
TechCrunch — Anthropic users face a new choice: opt out or share your data for AI training — independent reporting that the toggle defaults ON.