Privacy follows the authority boundary.

Starlight Intelligence privacy policy details

This policy covers starlightintelligence.ai, its public Department OS APIs and MCP endpoint, Memory Studio, the optional cloud curator, and the Starlight Growth Studio plugin. The operator is Frank Riemer / Starlight, Amsterdam. Questions may be sent to frank@frankx.ai.

What the site handles without an account

• Ordinary web request data may be processed by the hosting provider (Vercel) for delivery, abuse prevention, diagnostics, and security logging.
• Audience measurement uses Vercel Web Analytics: page views and a small set of named interactions (copying the start prompt or an install command, exporting Memory Studio files, and waitlist steps). These events carry the event name and, for copies, which host was selected. They do not carry the prompt, your files, or your memory content. Vercel Speed Insights also measures page performance, including Web Vitals, page path and browser/device information. Both integrations remove query strings and fragments before sending page URLs. There is no advertising tracker and no behavioral profile.
• The configurator sends only the five product-fit selections shown in its form. The response is deterministic and is not generated by a model.
• The public MCP endpoint processes the arguments required for its read-only catalog and planning tools. It applies bounded in-memory rate limiting using request network information.
• Memory Studio keeps your memory files, mission and editable brief in the current browser tab by default. Import and export happen as local file operations; nothing is synchronized to Starlight.
• If you choose device recovery, the workspace and any saved run reference are stored as plaintext in this browser on this device. Recovery is not isolated by Starlight account and survives sign-out. Anyone who can use this browser may be able to read it. You must explicitly restore saved work; you can remove the saved copy through the recovery controls. Keep confidential work out of shared browsers.
• The learning workspace keeps the entered brief and evidence draft in the browser. Export happens as a local download; the current implementation does not submit that draft to Starlight.
• Email messages are processed to answer the request and maintain necessary correspondence.

The optional cloud curator

The curator is off unless it has been activated for your membership. When you use it:

• Sign-in is handled by Clerk, which processes the account details you give it under its own terms. Starlight receives a verified identifier, not your password.
• Starlight stores, in its Supabase database (EU Central): the link between that identifier and your Starlight principal, your membership and entitlement status, a count of curator attempts per UTC day (five per day), and run records with a run reference, the model chosen, token usage, timestamps, duration, and status.
• The accounting run record does not store the mission you wrote, your memory files, or the generated brief.
• Your mission and memory files are sent through Vercel AI Gateway to the model provider you selected, for one generation. AI Gateway and the model provider process that content under their own terms and retention practices.
• When brief recovery is enabled, you may choose to keep a separate copy of the generated text for 24 hours. The choice starts off. The copy stores the original output, run reference, owner identifiers, timestamps and a hash used to match your locally saved source; it stores no mission or memory files. Only the original account in its active organization can retrieve it. Recovery reads stop at expiry. Active text is cleared through deletion or expiry maintenance; tombstones and accounting records remain. Database backups, downloaded files, browser recovery and model-provider retention have separate lifetimes. Retrieval and deletion remain available while generation or spending is disabled. The feature remains unavailable until its storage, consent and maintenance release gates pass.
• Checking a run receipt sends the saved run reference after sign-in and checks records belonging to your active membership. It returns lifecycle metadata, without retrieving generated text or retrying the model.

What the site does not do

There is no checkout, payment processing, managed mission runtime, cloud memory synchronization, or server-side Academy submission in this release. The Growth Studio plugin does not bundle, request, proxy, or resell model-provider credentials.

Repository improvement missions

The technical preview at /missions reads the public GitHub repository you name and sends your mission, selected source evidence and specialist outputs to OpenAI when you explicitly start a run. A session-only API key is sent over HTTPS through the Starlight server to the fixed OpenAI API endpoint. The application does not write that key to its database, browser storage, exports or application logs. OpenAI and hosting infrastructure process requests under their own terms; setting store:false does not replace their retention policies.

Mission state stays in this tab unless you choose device saving. Device saving uses plaintext IndexedDB in this browser profile, survives closing the tab and account sign-out, and can be read by people using the same profile. Browser clearing or eviction can remove it. Export your work for an independent copy. Repository reads use GitHub’s API; this preview supports public root Node packages with a recognized lockfile. Starlight does not execute repository code, install dependencies, push changes, deploy, publish or send messages during a mission. Exported diagnostic commands run only when you choose to execute them in your own environment.

Waitlists

Some pages carry a waitlist for a product that is not available yet. Joining is optional and works like this:

• What is stored. Your email address, the product you joined, your place in that list, the page you joined from, the referring site (origin only, never the full address), any utm campaign tags in the page link, and the time. If you answer the optional questions, your answers are stored with it: expected price band, role, how soon, what you are trying to do, and what you use today.
• Where. In a Redis database (Upstash, provisioned through Vercel) that only the site's server can reach. When a mailing list has been set up for that product, the address is also added as a contact to that product's own list on Resend, our email provider; otherwise it stays in the database only. Nothing is written to the site's source code or public files.
• What it is used for. Emails about that product only: when it opens, and when something real ships before then. The site does not send any email yet. Answers are read in aggregate to decide what to build first and what it should cost. Free-text answers are never published with your name or email attached.
• Abuse protection. To limit repeated submissions, the server counts requests per network address for ten minutes, using a keyed hash of the address. The address itself is not stored.
• Analytics. The form sends anonymous funnel events to Vercel Web Analytics (waitlist viewed, joined, answered, skipped) with the product and page path. They never include your email or your answers.
• Leaving. To have your record deleted from the database and, if you were added to one, the mailing list, email frank@frankx.ai. There is no automatic deletion yet.

Saved memory corrections

The Memory workspace offers a separately chosen private proposal inbox after account sign-in is configured. Creating a private workspace does not grant curator execution, funding or access to another client. Saving sends the correction's original words, separate interpretation, project, subject and supplied source references to the existing EU Supabase platform. These remain proposals; they do not change your canonical local memory or a signed creative edition. Source references are recorded as supplied and are not automatically verified.

Each record retains revisions and expires from retrieval after 90 days. A bounded maintenance operation erases expired active text and its stored revisions; its result is reported separately from other maintenance. The owner can explicitly grant expiring read access to a member of the same workspace, revoke that grant, export authorized corrections, or erase stored correction text and revisions. Erasure retains minimal references and request hashes for reconciliation. Prior exports, external copies and provider backups have their own retention; deleting active text does not erase those copies. Credentials and hidden reasoning must not be saved in corrections.

Your inputs and third-party runtimes

The plugin runs inside the runtime you choose. Codex, Claude, Grok, Antigravity, Hermes, OpenClaw, model providers, connectors, and any sources you open have their own data practices and terms. Store secrets only in the runtime's approved secret mechanism. Do not place confidential data, personal data, credentials, or private memory in a public package, prompt, log, evidence draft, or curator run.

Purpose, retention, and sharing

Request data is used to deliver and secure the service, compile the requested public plan, measure which pages and actions are used, run the curator when you ask for it, diagnose reproducible defects, and respond to communications. Starlight does not sell personal data. Hosting, analytics, identity, database, and model-routing providers process limited data as service providers. Account and usage records are kept while your membership is active and as needed for security and accounting; client-side content is not retained by Starlight unless you separately send it.

Your choices

You may use Memory Studio, the published skill, and the free tools without an account. You may ask about data associated with your account or a direct communication, request access, correction, or deletion where applicable, or raise a privacy concern by email. Applicable legal rights depend on your location and the nature of the request.

Changes

This policy will be updated before payments, managed runtimes, cloud memory, server-side learning submissions, or materially different analytics are introduced. The waitlist section was added on 5 October 2026; the effective date above identifies this version.