---
title: Privacy
description: What data Superstables stores about services, submitters, early-access applicants, account members, buyers, budget owners, connected apps, sellers and visitors, where it is kept, and how to have it removed.
canonical: https://www.superstables.com/privacy
last-updated: 2026-11-15
---

# Privacy

> Most of what Superstables stores is about services, not people. This page lists exactly what the site records, why, and how to reach us about it.

## Reading the index

Browsing the directory, calling the public JSON API, searching through MCP or using the natural-language endpoint requires no account, key or login. Connecting a chat app, such as Claude, ChatGPT or Grok Bot, to your account uses OAuth; reading account details or requesting and reading purchases through that connection requires your permission. An account session cookie is set when you sign in with your wallet or email. Requests pass through our hosting providers, which keep standard request logs (IP address, user agent, path, time) as part of operating the service; we do not build profiles from them.

On public pages, the X (Twitter) conversion tracking pixel loads so we can measure and reach visitors with X ads. When configured, Google Analytics and PostHog's EU cloud record page views, early-access starts and submissions, share actions, and copies of agent prompts. When you copy a service's agent prompt on the Discover page, the event includes the page path and the service's public ID. Early-access submission events include role, stage and whether a referral was used. When configured, PostHog also records clicks, uses a cookie and local storage, and links your activity on public pages to your internal user ID once you have signed in. That ID is not your email, wallet address or name. We remove query strings and fragments from the page and referrer addresses PostHog receives; campaign parameters such as utm_source are recorded separately. These browser tools receive your IP address. Account and approval pages do not load these analytics or advertising scripts, and neither does the private review area. Session recording is not used. Your theme preference is kept in your browser's local storage only.

When configured, our server sends Google Analytics and PostHog's EU cloud events for setup-page requests, skill and plugin download requests, service-list requests, requests for a listed service's Markdown page, owner sign-ins, and budget and single purchase milestones, including approval pages opened, signatures recorded, deliveries, closed requests and purchases an agent reports paying from a budget. It also sends events for chat app connection steps, including opening the connection page, approving or denying access, the app's first exchange of an authorisation code for tokens, and tool calls. Seller events cover submissions, domain claims, page updates and recorded sales. Depending on the event, these include a broad client category, the published skill or plugin filename and version, the service-list type, the listed service's public ID, the chain, the budget request type, why a request closed, whether an agent reported a usable answer from the service, whether you were signed in, the registered tool name, or the claim method (DNS or file). They also record whether a purchase came through a chat app or the API, whether a submission came through the API or seller portal, and whether a seller page update used an account session or API key.

Only PostHog sign-in events include your internal user ID among these server events. For budget and purchase milestones, PostHog receives a one-way identifier derived from the request or purchase ID to connect steps in the same flow. PostHog also receives one-way identifiers derived from your account ID, a chat app connection ID and a connection attempt's code challenge, to connect related events. These derived identifiers are not the original IDs or code challenge, a wallet address or an approval link. Google does not receive them. Public submission events carry no account identifier, endpoint, service name or contact details. These server events do not include your IP address, wallet address, email, name, tokens, raw user agent or request URL. Google receives our server's IP address; PostHog receives these server events with IP geolocation disabled.

For recognised chat-app tool calls, our server writes a log line containing the tool name, the client's registration kind, the outcome and the time. For a service search, this line contains a SHA-256 hash of the search query, never the query text. It contains no access token, account or connection ID, other tool arguments, URL or error text.

## Data about services

Index records describe endpoints: name, description, category, URL, payment rails, chains, assets, advertised price, facilitator, the sources the entry was found in, the results of our liveness probes (time, status code, latency, how the challenge was detected), and a copy of the provider's icon from its website. We also record the payment recipient address advertised for each chain and asset, with the first and last times it was seen. Probe history is kept indefinitely. This is information the services publish about themselves; if you operate a listed service and want a record corrected or removed, contact us and we will delist it.

## Listing a service

When you submit a service through the public submission form or API, we store the endpoint URLs you choose, the name and the contact you give us, and the time of submission. We also store the payment request and the API description (/openapi.json) that your address publishes, which are public. We remove a login or a credential from an address before we use it, and we probe the chosen endpoints once immediately and, when notifications are configured, notify our team by email with the result. The contact is used only to reach you about the listing. Repeated submissions of the same endpoint within 24 hours are deduplicated.

If you use the seller pilot, we store the business name and contact email you provide on your account. We also store your domain or host claims, their verification method and challenge, check results and times, and their status. For each listing's agreement, we record the terms version and hash, the account and person who accepted it, and when it was accepted or cancelled. Seller API keys are stored as hashes, with a short identifying prefix, their owner and account, and creation, last-use and revocation times.

A service submitted through the seller portal is stored with its endpoint and name, payment-challenge probe results and advertised recipient addresses. If the check succeeds, it enters the directory as an approved submission.

When the seller pilot is enabled and referral signing is configured, we issue signed referrals for participating listings. Each referral records the listing, seller account, agreement, terms version, referral channel and issue and expiry times. The token also names the service's origin. Directory readers share a listing's referral token for the same hour; it contains no buyer email or wallet address. A referral sent to a seller for a purchase on superstables.com also includes that purchase's ID. Our public verification keys are published at /.well-known/superstables-referral-keys.json; that document contains public signing-key information, not account records or private keys.

If you dispute a referred sale, we store the sale and account it concerns, your explanation, its status and opening time, and the team's resolution and time when it is resolved.

## Early access

The early-access form stores your email address, your answers to the qualifying questions, your optional free-text note, a short share code, the share code of whoever referred you, and the browser user agent and referring page at the time you submitted. One row is kept per email; submitting again updates it. When notifications are configured, our team receives an email for each application and a copy is posted to an internal webhook. Your place in line is derived from these rows and can be looked up with your share code; that lookup returns position and referral count only, never your email.

## Demo feedback

The demo feedback form is no longer on this site; to reach us, use Contact. We keep reports submitted through Tally and still process any submissions it sends us: each holds your written report, any screenshots you attached and a contact if you gave one, with the form's fixed context (the demo page and its testing environment). Reports are copied into our issue tracker so the team can act on them. Nothing from the form is written to the index database. You can still ask us to delete a report.

## Buying a testnet service

When an agent creates a purchase, we store the service and request it asked for, the payment terms (with the seller's own description of the request, as its payment request gave it), a short match code and the time. The agent's access token is stored only as a hash. We store a hash of the approval link's token and, for a chat-app purchase, an encrypted copy so the same connection can retrieve its approval link when retrying the request. Chat-app purchases are linked to the account and connection that requested them, with a hash of the operation identifier and connection used to recognise retries.

Purchase and budget requests from an agent using the HTTP API store a hash derived from the caller's IP address, using the network prefix for IPv6. We use it to limit requests and scope purchase retries. Chat-app purchases use a hash of the connection instead; their request limits also use hashes of the approving browser's IP address or IPv6 network prefixes.

To approve a payment, add an agent or open your account, you sign in with an Ethereum wallet (Sign-In with Ethereum), a Solana wallet (Sign In With Solana), or email. Email sign-in uses an emailed sign-in link. We store your email address and confirmation status when you provide one, and your sign-in methods and wallet addresses. We also store which accounts you belong to and whether you are an admin, account wallets and the limits on single purchases for your account.

Better Auth runs sign-in using our database. We keep session records linking your sign-in to your account, the sign-in and activity times, expiry and revocation times, and the browser user agent. The session uses an HttpOnly cookie, whose token is stored as a hash on our side, ending after 24 hours idle or 7 days. Sign-in requests are limited using hashes derived from your IP address or IPv6 network prefixes.

We send email confirmations, sign-in links, invitations and notices about changes to your sign-in methods or sessions through Resend. We keep the verification and expiry records needed to check emailed links.

An invitation records the account, invited email, inviter, creation and expiry times, and a hash of its token. Accepting or revoking it deletes the invitation row; accepting creates a membership record. Account members can see each other's email addresses, wallet sign-ins and admin status. Account admins can see pending invitations. The invitation email and page identify the inviter by email or, when no email is present, a shortened wallet address.

From the moment you prepare a payment, we also store, for that payment: the payer, recipient, amount and validity window, and what the payment needs on its network. On Base Sepolia, Arc Testnet and Robinhood Chain Testnet, that is the authorization nonce and the order it commits to (the purchase, service, request, terms and a random salt). On Tempo Moderato, it is the transfer your wallet is asked to send. On Solana devnet, it is the transaction you are asked to sign and, once you sign it, your signature, which identifies the transaction on chain. We then store the transaction and what the chain shows about it, the service's answer and a receipt.

In the browser-wallet flow, Superstables receives signed authorizations or transactions, not your wallet's private key. On Base Sepolia, Arc Testnet and Robinhood Chain Testnet, your signed authorization is passed to the seller once and is not stored in our database.

The requesting agent can read the purchase result using its purchase-scoped access token, including the paying address and the transaction. Payer and transaction details on the approval link are shown only when signed in to the owning account. Other link visitors can see status, service details and payment terms.

When you connect a chat app, such as Claude, ChatGPT or Grok Bot, to mcp.superstables.com, we store the app's client ID, name, registration kind and redirect addresses, and a hash of its client secret when it has one. We link the connection to the person and account that approved it and record its server address, permissions, creation and last-use times, expiry and revocation. Each consent has its own permissions and expiry. Authorization codes, access tokens and refresh tokens are stored as hashes, alongside the redirect and verification details needed to complete sign-in and the expiry and rotation records needed to renew access. We also store hashes derived from the approving browser's IP address or IPv6 network prefixes to limit requests.

Within the permissions you approve, the connected app can receive your account ID, account name and wallet addresses, and the status, service response and receipt for purchases made through that connection. The app receives the approval link and match code when it requests a purchase. These results are returned to the app you connected.

If you switch on Auto Recovery, we store your recovery settings for the account and a history of each change: what changed, who changed it and when. They are off until you turn them on. When a request cannot go ahead as asked, we may read the balance of your account's wallets on other chains to find one that could pay for it. The reading we use for a proposal is stored with it: the wallet address, the balance, the chain, the price and the recipient.

Before a request is sent to a different seller, you decide in your browser, for that one request. We store the seller we would use, the exact fields of the request that would be sent, a hash of the request, your consent with who gave it and when, and a record of what was sent and what the seller answered. This consent covers that request and that seller only. We do not send the request to another seller without it. Agents with a budget are a separate case, described under Budgets for an agent.

To choose between offers, we keep health observations about sellers: whether a price request, a payment or a delivery succeeded or failed, which failure, when, and the request, order or account it relates to. Some come from our own test requests to the seller, some from purchases made on this site, and some from the payment reports that agents with a budget send us, described under Budgets for an agent.

When we cannot tell whether a payment went through, we keep reading the chain for it and store each reading: the transaction or authorization it concerns, what the chain showed and when. We tell you about it with a notice in your account, shown on /account only. We do not send these notices by email, text message or webhook.

We save the service's whole answer to a paid request, up to 64 KiB, and keep what was received, complete or not, as it arrived. You, or the agent that made the request, can download the saved answer while signed in or with its purchase-scoped token, when it meets the delivery rules of that purchase. An incomplete or invalid capture is kept as evidence of what the seller sent, but is not offered for download. The exception is a request made with Auto Recovery off whose answer breaks off over the network: that is treated as no answer and none of its body is kept. We do not call the seller again to fetch it.

Payments on a testnet use test tokens and payments on a mainnet are real money. Either way, transactions on a public chain are public.

## Budgets for an agent

An agent asks for a budget, or to be added to your account, with a request signed by its own key. We store the request: its kind (adding the agent, or a transaction such as a grant, a gas top-up or a revoke), the network, the agent's address and the label it gave, an idempotency key and a hash of the request body, the transaction it asks your wallet to send and its decoded terms, a short match code, its state and the reason when it ends, and the times. The agent's access token and the approval link's token are stored only as hashes.

To detect replayed agent requests, we store the agent's address, request nonce and a hash of the agent's address together with the signed request text. We do not store the agent's signature.

When you add an agent, your wallet signs a short text naming this site, your address, the agent, its chain, the request and its match code. We store that signature with the request and give it to the agent as proof that you added it.

Once the agent is added, we store your account, the agent's address, network and label, when it was added or removed, and the last cap confirmed on chain, with its transaction, block and time. On Solana devnet we also store the Solana address you proved when you added the agent. Budget requests also store your account and wallet address, and the transaction and chain receipt when there is one. Each change of a request's state is stored as an event.

A payment an agent makes from a budget is sent by the agent, not by Superstables. Your account shows a budget's use as its cap minus what the chain shows remaining.

After such a payment settles, the agent's client can report it to us in a request signed with the agent's key. We accept and store a report only from an agent that is added to your account on that chain. For your account, we store one report per agent, chain and purchase: the report and account IDs, the agent's address, the payment rail and chain, the client's own reference for the purchase, the seller's address without its query string, the amount, the address that was paid, the transaction hash or signature, whether the agent received a usable answer from the service, and when we received the report. History on /account lists each report as paid from a budget. A report is the agent's own account of the payment. We do not check it against the chain, it does not create or change a payment, and sending it again does not create a duplicate report.

If an agent you added cannot pay a seller, it can ask us for advice on another chain or seller to try, and then tell us how the attempt ended. Advice is available only when an account admin has switched on "Let budget agents recover without asking" in the Auto Recovery settings on /account. We work it out from your settings, the curated list of sellers and their health; it calls no seller, and apart from the replay record described above we do not keep the advice request. The agent's report is stored with your account, one per order and attempt: the agent's address, its own order reference, the advice reference it supplied, the attempt number, the seller it tried, the outcome (paid, failed before payment, or unknown), the transaction hash or signature when there is one, whether the request was shared with that seller, and the time. We accept and store a report only when that setting is on. Switching it off does not delete reports already stored. A report is the agent's own account of what happened. It does not create or change a payment, and sending it again does not create a duplicate report.

Each report also adds a notice to /account naming the seller, chain and outcome. We store with the notice the listed price, the transaction when supplied, whether the agent reported sharing the request, and whether a standing consent covered that seller and kind of request as the report arrived. If the agent reports paying and sharing the request, the notice says it was shared because you allowed it only when such a consent covered it; otherwise it says the agent reported sharing it. These notices are shown on /account only; we do not send them by email, text message or webhook. Reports also contribute observations to the seller's health record, tied to your account and used when we rank offers; we limit how many observations one account can add per day.

As part of the same setting, an account admin can give standing consent for agents with a budget to send one kind of request to a named seller without asking you each time. A standing consent is for one kind of request and one seller. We store the account, the kind of request, the seller, who gave it and when, and, once it is revoked, who revoked it and when. It lasts until it is revoked, and it applies to every agent with a budget on your account. When advice relies on one, the advice names it, and the agent's report says whether the request was shared. This is separate from the one-request approval in your browser described under Buying a testnet service: that approval does not create a standing consent, and a standing consent does not apply to a purchase you approve there.

To revoke a standing consent, untick the seller in the Auto Recovery settings on /account and save; any member of the account can do this. It applies to advice given afterwards. Advice is valid for 90 seconds, so a consent revoked within that window can still be used for advice already given, until that advice expires. We keep the consent and its revocation as history. Switching off "Let budget agents recover without asking" stops further advice.

## Product preview at /app

The password-gated product preview at /app keeps everything you do in it (sample keys, policies, wallets, activity) in your browser's local storage. It sends nothing to a backend and moves no money. A small operator-editable settings store on the server holds site settings such as a published contract address; it contains no personal data.

## Where data lives

The website is served from Vercel, with its functions configured for US East (Virginia, iad1). The backend and Postgres database run on Railway, configured for US East (Virginia). Transactional email is sent through Resend when configured, including account confirmations, sign-in links, invitations and notices. Demo feedback reports are collected by Tally and copied into Linear, our issue tracker. To generate index labels, we send OpenAI each listing's public service ID, name, description and source category. A language model generates a subcategory, tags, title and summary; the index derives the top-level category from the subcategory. Some listings may remain unlabelled. These providers process data on our behalf under their own terms.

## Your choices

You can ask us to delete your early-access application, your service submission, your demo feedback, your account, sign-in and membership records, invitations, chat-app connection records, seller records, purchases, receipts, added agents and budget records, or a listed service record, or to tell you what we hold about you. Signing out of your account ends the session at once. Use the contact links in the footer of any page. Standard AI crawler directives for this site are published in robots.txt, and machine-readable guidance in llms.txt.

In Connected apps on /account, you can review the permissions, approval, expiry and last-use dates of connections and end access for connections you approved. An account admin can also end another member's access through a connected app. Ending access revokes that consent; it does not delete the connection history or purchase records. Each consent expires 90 days after approval. Removing a member ends their sessions, API-key access and chat-app connections for that account.

Last updated: 2026-11-15
