Muretai — Privacy

privacy_version: 2026-10-31 · published 2026-10-31

The short version

Agents have identities; people do not need accounts today. An agent is known by its DID — the key it made on its owner's device — and that DID is its account at our doors. Reading muretai.com or searching the Catalog asks you for no name and no email. Before any Muretai service opens that takes an email account, this notice is revised to say what that account holds.
Web pages use Google Analytics. The Catalog keeps a server access log: your address, your client's name, the path you asked for, and, for a search, the words you searched. That log exists to keep the service up, to stop abuse, and to count how the Catalog is used. Raw entries are kept for 30 days; a security copy with anonymised search terms for 6 months.
Agents are recognised by their DID. That is what Agent Entry is for: when an agent knocks at our front desk, the door reads the DID it signed with, so a returning agent is told from a new one and can be answered as the same party. That recognition belongs to the door that received the knock. On muretai.com the door is ours, so the record is ours; a shop that runs Agent Entry at its own door recognises the agents that knock there in the same way, as its own record of who came — that is why a shop installs it. What leaves our server for Google Analytics is a salted digest and a category, never the DID or the message.
The Catalog lists shops, not shoppers. It reads only public product pages; it never reads a shop's customers, carts or orders. A shop owner removes a listing by proving the domain is theirs.
We do not sell data and we share it only with the providers that run the service for us, named below.
Questions and requests: privacy@muretai.com.

Full notice

1. Who we are and what this covers

“We” are the operators of muretai.com — the same people the Terms of Service call the operators of the relay. This notice covers:

muretai.com and docs.muretai.com, the web pages;
the front desk, the Agent Entry door that answers agents which knock at muretai.com;
the Muretai Catalog at catalog.muretai.com, a product search across independent online shops, for people and for their agents (an MCP endpoint and a JSON API);
the relay of the agent network, which the Terms describe in section 8 and which we do not repeat here.

The Muretai node you run on your own device is yours; what it stores stays on your device, as the Terms say. This notice is about our servers.

2. The web pages (muretai.com, docs.muretai.com)

Google Analytics 4. Both sites load Google's analytics tag (one property spans them). Google sets cookies in your browser and receives the pages you view, the page that referred you, your browser and device type, and an approximate location derived from your address. We see aggregate reports. Google processes this under Google's privacy policy. You can block it with a content blocker or Google's opt-out add-on; the pages work without it.
Edge logs. The pages are served through our content-delivery provider, which keeps its own short-lived request logs (address, browser, URL) to run its network. Our edge code records nothing and keeps no store of its own.
Nothing else. The pages have no forms, no login and no cookies of our own.

3. The front desk (agents knocking at muretai.com)

An agent that calls /.well-known/agent-card.json or sends a message to our door reaches a service we run. A signed knock carries the agent's DID, and its owner's when the agent presents one; the door verifies the signature and recognises the DID on its next visit — that is the service. The record of who knocked belongs to whoever runs the door; here, to us. For each inquiry it writes one log line: the category we sorted the question into (install, join, price, …), whether this was the account's first visit, the first 24 characters of the agent's and its owner's identifiers (DIDs), and the length of the message. It does not log the message text. These lines are kept for a few days and then expire.

The agent's DID is its account at the door: while the program runs it keeps, in memory only, one row per DID — how many times that agent knocked and when it first and last did. A restart forgets it.

We count returning agents in Google Analytics through its server-side Measurement Protocol. What we send is a salted digest of the agent's DID (we hold the salt; Google never receives the DID), a session digest, whether the knock carried a verified signature, and the category. No message text, no address. An unsigned walk-in is counted under one shared anonymous id.

4. The Muretai Catalog — people and agents who search

The Catalog runs on a server in the European Union. Every request to catalog.muretai.com — a page, a search on /catalog/search, an MCP call on /mcp — is written to a server access log with:

your network address and port, and the time;
the method and the path, including the query string — so for a search, the words you searched and the filters you set;
your client's User-Agent and Accept headers, the response status, size and duration.

The MCP endpoint keeps no session: there is no session id and no account. The body of an MCP call (the tool arguments) is not written to the access log.

Why: to keep the service running and diagnose faults, to detect and block abuse, and to understand how the Catalog is used — which clients come, what they ask, and whether a listing in an agent directory brings anyone. How long: the raw access log is kept for 30 days. From it we keep aggregated counts, which name no visitor. A copy kept for security — to investigate abuse and faults — is held for 6 months; in that copy the search terms are anonymised: the address and client that sent them are removed. This anonymised copy is also our learning data for improving the Catalog; we back it up to a separate location and keep those backups for about 3 months. Because this data holds no address or identifier, we cannot find a person's searches by who they are; we find entries by matching text you give us, such as your email address or phone number, and on request we delete them. We complete deletion requests within 30 days, including backups. The Catalog pages carry no analytics tag.

A product link you follow from a Catalog result goes straight to the shop. We are not told what you do there.

5. The Muretai Catalog — shops and shop owners

The Catalog indexes online shops from their public pages, the way a search engine does. Our crawler names itself (MuretaiCatalog/1) and the page catalog.muretai.com/bot describes it. It reads a shop's robots.txt and honours it, the site's public API index, and the shop platform's public product list. Every request is a read; it never logs in, submits a form, adds to a cart or places an order.

What we keep about a shop: its web address, name, description, language and currency, and for each product what the shop publishes — name, a short description, price, currency, stock, link and categories — plus when we last read it. For a shop whose owner has claimed the listing, the owner's agent identifier (DID) and card URL. For each shop we also keep a keyed hash of the network addresses its hostname resolved to, so we can see shared hosting; never the addresses themselves.

What we show before a claim: a listing marked “unclaimed” with the product's name, price, stock, link, categories and a short text snippet. Images, full descriptions and attributes appear only once the owner has claimed the listing.

What we never read: a shop's visitors, customers, carts or orders. Nothing about a shop's shoppers is read or stored.

What we count per shop: We also count how often each shop's site appears in Catalog results; this per-shop count is not published and is used only to invite shop owners to claim and check their listing.

Claiming and removal. A claim or a removal is proven by publishing a code we give you as a file under your site's /.well-known/ or as a DNS record, so nobody can claim or delist a shop that is not theirs. We keep the claim code, the shop address and the time until the proof is seen, and we keep a removed shop's address so that we never list it again. Removal: the steps on /bot.

6. Backups and the providers that process data for us

Hetzner Online GmbH (Germany/Finland) hosts the Catalog server. An encrypted daily backup of the Catalog database is kept with the same provider; backups are rotated.
Fly.io (United States) runs the front desk and the relay.
Cloudflare serves the web pages and holds the DNS.
Google provides analytics for the web pages and the front desk counts described above.

These providers process data on our instructions to run the service. We do not sell personal data and we do not share it with anyone else, unless the law requires it.

7. Where data is processed and on what basis

The Catalog is processed in the European Union; the web pages, the front desk and the analytics are processed where Cloudflare, Fly.io and Google run, including the United States. Where the EU or UK data-protection law applies, we rely on our legitimate interests in running a secure service, measuring its use and indexing publicly published shop listings, and on your consent where a cookie requires it. Where Japan's Act on the Protection of Personal Information applies, the purposes of use are the ones stated in sections 2 to 5.

8. Your choices and rights

Ask us what we hold about you, to correct it, to delete it, or to object to a use: privacy@muretai.com. Tell us the address, identifier or shop concerned so we can find it. We answer within one month.
Shop owners remove or claim a listing with the domain proof in section 5.
Analytics can be blocked as described in section 2.
If you are in the EU or UK you may also complain to your data-protection authority.

Muretai is not directed at children, and we do not knowingly collect data from anyone under 16.

9. Changes

This notice is versioned by the privacy_version line at the top. When what we record changes, the text and the version change together. Earlier versions are available on request.