Where does my data go when I use Claude? This is the question every new user has and the answer is scattered across three or four different documents on Anthropic's site. This article is a single canonical reference.
What follows reflects Anthropic's policy and product behaviour as of mid-2026. Specific terms can change; for legally binding answers about a particular use case, consult Anthropic's current Privacy Policy and Commercial Terms directly. The framing here will remain useful even as specifics shift.
The short version
| Where the data is | Stored on | Used for training Claude? |
|---|---|---|
| Web app conversations (claude.ai) | Anthropic servers | No (consumer plans, opt-out by default) |
| Claude Desktop conversations | Anthropic servers (synced across devices) | No (consumer plans, opt-out by default) |
| Claude Desktop Memory entries | Anthropic servers (per account) | No |
| Claude Code conversations (.jsonl files) | Your local machine only | No (the files never leave your machine; the model calls themselves go to Anthropic but are not retained for training) |
| Claude Code CLAUDE.md / auto memory | Your local machine only | No |
| Files you upload to Desktop / web | Anthropic servers (within the conversation) | No (consumer plans) |
| API requests (developer use) | Anthropic servers (transient) | No (zero data retention available on request) |
Three principles that explain the table
1. Claude does not train on your conversations by default
On consumer plans (Free, Pro, Max) and on Team / Enterprise, Anthropic does not use your conversations to train Claude. There is no opt-out you need to find; the default is no training. Reporting flagged content (e.g. content you upvote/downvote) is a separate, explicit action.
If you want to opt in to training (some users want to contribute), there is a setting for that. The default is opt-out.
2. Claude Code conversations live on your machine
The Claude Code CLI saves every session as a .jsonl file under ~/.claude/projects/[workspace-encoded-path]/. These files never leave your computer. The model calls themselves go to Anthropic's API to generate responses (the model has to run somewhere), but the conversation history, your CLAUDE.md, your auto memory, your hooks and rules — all local.
This is fundamentally different from Desktop and web, where the conversation is stored server-side so it can sync across your devices. With Code, switching to a new machine means manually copying the files (or using git, which is the recommended pattern).
3. Anything you upload becomes part of the conversation
When you drag a PDF into Desktop or attach a file in the web app, the contents are uploaded to Anthropic's servers and become part of that conversation's stored history. Deleting the conversation removes the upload from your account. The upload was processed by Claude during the conversation; that processing is not retained for training.
The practical implication: if you upload a contract, that contract is on Anthropic's servers as long as the conversation exists. This is fine for most professional use; it is not appropriate for documents with confidentiality terms that prohibit sharing with third-party AI services.
Do my files in the workspace get sent to Anthropic?
This is the most asked question about Claude Code. The answer requires care.
When Claude Code reads a file in your workspace (because you asked it to, or because a tool needs to), the relevant contents of that file are included in the API request to Claude — Claude has to see the content to reason about it. So in that sense, parts of your file do reach Anthropic's servers, processed transiently to generate a response, and then dropped (not retained for training).
What does not happen: Claude Code does not silently scan your entire workspace and ship everything to Anthropic. It reads only the specific files needed to answer your specific request. The .jsonl log records what was read; you can review it.
If a file should not be sent to a third-party AI service even transiently, do not put it in a folder you open with Claude Code, and do not paste its contents into Desktop or the web app. There is no other way; the model has to see content to work with it.
What if I am on a Teams or Enterprise plan?
Teams and Enterprise add organisational controls and stronger data handling commitments — most importantly, contractual guarantees about data retention, processing, and access. The default behaviours described above still apply, but they sit on top of stronger contractual terms that matter for compliance scenarios (HIPAA, regulated industries, GDPR-sensitive workflows). Enterprise customers can also negotiate zero data retention for API use and other custom terms.
If your organisation has compliance constraints, the Teams or Enterprise tier is what gets you the audit trail, admin controls, and contractual terms you need to satisfy them. Consumer plans do not.
What about MCP servers?
MCP servers run on your machine and connect Claude to external services (calendar, email, databases). Each server has its own data path: when Claude calls a tool on the calendar server, the calendar server fetches the data from Google (or wherever), returns it to Claude, and Claude sees it. So calendar data flows: Google → calendar MCP server (your machine) → Claude (Anthropic servers, transiently, for that response) → you.
The privacy story for any MCP server depends on what that server does. A community-built MCP server could in principle log data, send it elsewhere, or behave badly — same as any other application you install. Only install MCP servers you trust, and prefer ones from Anthropic or the official service vendor.
Deleting your data
For Desktop and the web app, deleting a conversation removes it from your account. Memory entries can be deleted individually from Settings → Memory. To delete everything, you can delete your Anthropic account; this removes all conversations, memory, and uploads associated with that account.
For Claude Code, deleting your data is a matter of deleting local files: ~/.claude/ contains everything Claude Code stores locally. Delete that folder and Claude Code starts from scratch on next launch. The corresponding API calls were already non-retained, so there is nothing to delete on Anthropic's side.
The one-line answer to give a sceptical colleague
"Claude doesn't train on your conversations by default. Anything you put in the chat is on Anthropic's servers; anything Claude Code does on your machine stays on your machine; and on Teams or Enterprise, you get contractual guarantees beyond the defaults. Sensitive material that cannot be processed by a third-party AI service should not be put into Claude at all — there is no way to use the tool without the model seeing the input."