Privacy and data use

Lithi uses the data categories your product, task, and connection choices put in scope. For Compute, you choose the purpose and permitted data profile. The Kit prepares and encrypts permitted input before transfer, while an assigned Mac can read its piece during execution.

Start with Choose a Compute data profile when you need to decide whether an input fits.

The data categories that matter

The relevant category depends on what you ask Lithi to do:

• Account and contact details: name, email address, organization details, and information needed to administer access or answer a request.
• Team and workspace records: members, roles, settings, and other records needed to operate a shared workspace.
• Service and connection records: the selected service, account or resource, permission state, configuration, and visible connection activity.
• Work context and outputs: approved references, instructions, task results, and records your workflow creates or asks Lithi to return.
• Support and security details: the information you choose to include when asking for help or reporting a concern.

These categories are a practical guide, not an exhaustive field list. The purpose, selected scope, and terms that govern your account determine what is relevant.

Provider operational telemetry is a separate category

Provider telemetry describes a participating Mac's runtime and resource use, separately from customer task content. It can include permitted coarse region evidence for impact reporting. It does not grant access to task inputs or outputs.

Read data and processing boundaries for the permitted fields and retention rules, and Compute impact methodology for how evidence and missing coverage appear in a report.

How an AI-assisted task stays scoped

The data boundary for AI-assisted work depends on the product, task, selected sources, permissions, service path, and current commercial terms. Running an original task does not by itself authorize indefinite retention, an exact-copy learning dataset, or model training. A secondary use needs its own stated purpose, permitted scope, and applicable policy or agreement; some workflows may permit it and others may not. Read the retention guidance and data and processing boundaries before relying on a broader statement about training, retention, logging, or deletion.

Keep approved sources, account records, connection data, and operational records as separate categories. A permission to use one source for one purpose should not be treated as permission to use every source for every purpose. Ask for the current contract and provider record when a decision depends on model training, prompt retention, diagnostic logging, human access, regional processing, or deletion.

Know the four roles and what each can see

Four roles can see different parts of a Compute task. You decide the outcome, the data, and the policy. Lithi plans, admits, and checks the work. An approved provider Mac reads only the task content admitted to it. A customer-chosen AI client is a separate, fourth path with its own permissions and terms — connecting one does not change what Lithi or a provider Mac can see.

Local execution is not automatically local disclosure

Running a command in your own environment is not the same as keeping its data local. A terminal agent can send a local command's output to its own cloud model. A notebook can log rows to a remote tracker. An editor can index your files into its context.

That is why preparation tools return safe summaries and handles by default, and keep raw values and identity maps out of ordinary agent context. A project ignore file reduces accidental inclusion; it is not an access control. The paper explains the same boundary in Prepare and protect your input and Split work without losing meaning or privacy.

Secret scanning and personal-data detection are separate checks. Neither substitutes for the other, and a clean run of one says nothing about the other. See what leaves your machine for the three decisions a preparation run can reach before anything is sent.

When a provider is connected

Connecting a provider can change the path for selected records. You choose the service, account, resource, purpose, and scope; data movement depends on that choice and on the selected service's handling. A connection does not authorize unrelated accounts or resources.

Review the permissions shown during setup and remove a connection when its purpose ends. For the related service boundary, read Compute security and the subprocessor overview.

Website chat is a separate path

The public Ask Lithi chat handles the conversation you choose to send there. It uses the website's chat provider and an operational conversation record to answer and continue that chat; it is separate from customer Compute execution. A customer-chosen AI client is also a separate path with its own permissions and terms. Do not paste private Compute input into either path unless you intend to share it there.

How Ask Lithi uses your visit. When Visit context is allowed, Ask Lithi uses context from your current visit to tailor its answers. The site records which pages you open, how long each page is visible and in active use, how far you scroll, and which page sections appear on screen and for how long. It uses these observations only during your visit to suggest questions, put the most relevant answers first, and choose what to explain and ask next.

The relevant page context, including these measurements and the published text of the sections you view, goes to Anthropic with your question. Ask Lithi may mention a page or topic you looked at, but it does not list your visit back to you or show times or scroll depth. These measurements do not establish your identity, actual reading, or comprehension. Lithi does not use this context to identify you, build a cross-visitor profile, or train model weights.

Visit context follows the visitor's region. In the EU, EEA, UK and Switzerland, it stays off until you allow it in Privacy choices; elsewhere, it is on by default unless you turn it off there. A saved personalization opt-out remains respected, and an invalid, expired, or ambiguous saved choice keeps Visit context off. Global Privacy Control keeps Visit context and advertising measurement off for this browser. Advertising measurement is separate and stays off until you enable it in Privacy choices in the footer; you can change or withdraw it there, and changing advertising does not change Visit context. Page observations are deleted after 90 days and page views after 180 days. The chat still works without Visit context. For the length of your visit, the site also remembers which page and campaign link brought you here, so a sign-up records where it came from. This follows the same Visit context setting and never includes advertising click IDs.

Some processing supports the business around your task

Some processing exists to deliver the product you asked for — the task itself. Other processing exists to run the business around it: creating your account, sending an invoice, verifying a sign-in, or answering a support message.

These relationships have different purposes and terms. The service provider register lists each provider's purpose, data categories and processing locations. It separates Compute infrastructure from account email, support, website chat and measurement. The Privacy Policy and your agreement explain the terms that apply to those activities.

Provider roles follow the processing performed and the applicable terms. A service you choose independently, such as your own AI host, follows that separate relationship.

Make a privacy request

You can ask about access, correction, or deletion of information. Include the request type, your account or organization context, and enough detail to identify the relevant information or activity — for example, the affected workspace, date range, or record reference. To make a request, email privacy@lithi.ai.

Lithi confirms your authority over the account before acting on the request, and may ask you to clarify its scope. Send only the minimum necessary context: never include passwords, access codes, payment credentials, private message bodies, or unrelated personal information in your first message.

For a delivered Compute result, Read and verify Compute result files explains what the descriptor, transfer grant, byte checks, and destination read-back do and do not prove. For key custody, the paper's Choose who holds the keys section explains the difference between managed and customer-held custody.

Read the governing terms

This page is an orientation. The Privacy Policy and Data Processing Agreement provide the governing legal text for the terms that apply to your account.