Skip to main content
Your sessions land in a project that only you own. An administrator reads them only through an audited path that shows a banner for the whole visit. And you can drop the conversation content at the organization level, before it is ever stored.

Where your data lands

By default, your sessions land in your personal project: one project, one owner, you. Other members’ queries never touch it, because it is a separate project rather than a filter on a shared table. A session lands in a shared project only when you, or an administrator setting up a machine, pinned the tool to one with --project. See Team projects.

What a teammate can see

None of your personal project is shared. Teammates see your sessions only in a team project the tool was deliberately pinned to, under that project’s access rules.

What an administrator can see

Spend and usage always aggregate: the governance dashboard shows cost per user, per team and per department, because cost accounting is the point of the product. The aggregation holds numbers only; conversation content is not in it. Reading your actual sessions is possible for an organization administrator, and it is deliberately never silent:
  • The only path is the drill-in from the governance dashboard, which shows a persistent “Viewing as admin” banner for the whole time. The banner is enforced on the server, not painted by the browser.
  • Every such read writes an audit-log row. You can read the log yourself at Settings → Audit log and see every administrator read of your sessions.
  • There is no bypass flag and no service-token side door around the audit hook.
Permissions gate what any viewer sees even inside an allowed project: a viewer without content permission gets session rows with the generated titles blanked, and a viewer without the cost permission sees no cost at all rather than a zero.

What the organization can remove

An organization that must not store conversational content sets input, output and system instructions to Drop in the data privacy policy. LangWatch strips dropped content at ingestion, before storage, on every ingestion path, coding-agent telemetry included. What survives is metadata: model, tokens, cost, latency, timings, status, identity. The rollups keep working, and no person can read what was asked or answered. The named posture for this is no-spy mode. Scope it to personal projects only to cover the coding assistants without touching product observability. If the requirement is “only a small audience may read it” rather than “it must not exist”, Restrict keeps the content for a chosen audience and shows everyone else a [REDACTED] placeholder with the reason.

What you control from your own machine

You decide content capture at the source, before anything is sent. Each agent’s setup page names its flags; for example, removing the OTEL_LOG_* flags from Claude Code’s settings keeps the usage numbers flowing and stops the content. See the Content and privacy section of your agent’s page.

Secrets

Credentials that appear in prompts or tool output are replaced with a [SECRET] marker at ingestion, on by default for every project. Secrets redaction covers how detection works and how to extend it.

Retention

Stored data expires on a schedule the organization sets, resolved per project. See Data retention. Also check: Data privacy policy for the full policy reference: scopes, dispositions, audiences and PII redaction.
Last modified on August 27, 2026