The Extensible Messaging and Presence Protocol (XMPP) as defined in &rfc3920; does not support direct client-to-client interactions, since it requires authentication with a server: an XMPP client is allowed access to the network only after it has authenticated with a server, and the server will not grant access if authentication fails for any reason. If an unauthenticated client attempts to communicate directly with another client, such communication will fail because all XMPP communications are sent through one or more servers and a client cannot inject messages onto the network unless it first authenticates with a server.

However, it is possible to establish an XMPP-like communication system on a local (or even wide-area) network using the principles of zero-configuration networking. In this situation, the clients obviate the XMPP requirement for authentication with a server by relying on zero-configuration networking to establish serverless communication using the _presence._tcp DNS SRV service type. Once discovery has been completed, the clients are able to negotiate an XML stream between themselves (as generalized in &xep0246;) and then exchange messages and other structured data using the XMPP &MESSAGE; and &IQ; stanzas.

Serverless messaging is typically restricted to a local network (or ad-hoc wide-area network) because of how zero-configuration networking works. It is impossible for clients that communicate via this serverless mode to insert messages into an XMPP network, which is why this kind of "mesh" is most accurately referred to as an XMPP-like system that exists outside the context of existing XMPP networks (though see the Security Considerations regarding the ability to "forward" messages from a serverless mesh to an XMPP network or vice-versa).

Such a "mesh" can be quite valuable in certain circumstances. For instance, participants in a trade show or conference, users of the same wifi hotspot, or employees on the same local area network can communicate without the need for a pre-configured server. For this reason, support for serverless messaging has been a feature of Apple's iChat client when operating in Bonjour (formerly Rendezvous) mode for many years. Because it is desirable for other Jabber/XMPP clients to support such functionality, this document describes how to use zero-configuration networking as the basis for serverless communication, mainly for use on local links (although the protocol can also be used on ad-hoc wide-area networks).

This section provides a friendly introduction to serverless messaging. The examples show usage on a local link using dynamically configured link-local addresses as described in &rfc3927; (see the Wide-Area Networks section of this document regarding non-local usage).

Imagine that you are a Shakespearean character named Juliet. You are are using your laptop computer (a machine named "pronto") at a wifi hotspot in downtown Verona and you want to find other people to chat with on an ad-hoc basis (i.e., not people in your normal XMPP roster). Therefore your chat client advertises a serverless address of "juliet@pronto" so that other people can dynamically find you at the hotspot. Your client does this by invoking a daemon on your machine that supports DNS-based Service Discovery ("DNS-SD") as defined in &dnssd; and Multicast DNS ("mDNS") as defined in &mdns;. As a result, the daemon stores the following DNS records and listens for multicast DNS queries asking for them:

The meaning of these records is as follows:

• The A record specifies the IP address 10.2.1.187 at which the "pronto" machine will listen for connections.
• The SRV record (see &rfc2782;) maps the presence service instance "juliet@pronto" to the machine "pronto.local." on port 5562.
• The PTR ("pointer") record (see &rfc2317; and &rfc1886;) says that there is a service of type "presence" on the local subnet (".local.") called "juliet@pronto" and that the service communicates over TCP.

Your chat client also wants to advertise some information about you (subject to your control so that you don't divulge private information). Therefore it invokes the mDNS daemon to also store a single DNS TXT record (see &rfc1464;) that encapsulates some strings of information, where the record name is the same as the SRV record and the record value follows the format defined in draft-cheshire-dnsext-dns-sd (note: the line breaks are provided only for the sake of readability).

Other people at the hotspot can also advertise similar DNS records for use on the local link. Essentially, the mDNS daemons running on all of the machines at the hotspot collectively manage the ".local." domain, which has meaning only at the hotspot (not across the broader Internet). Queries and responses for services on the local link occur via multicast DNS over UDP port 5353 instead of via normal DNS unicast over UDP port 53. When a new machine joins the local link, it can send out queries for any number of service types, to which the other machines will reply. For the purpose of serverless messaging we are interested only in the "presence" service, but many other services could exist on the local link (see dns-sd.org for a complete list).

Now let us imagine that a fine young gentleman named Romeo joins the hotspot and that his chat client (actually his mDNS daemon) sends out multicast DNS queries for services of type "presence". To do this, his client essentially reverses the order of DNS record publication (explained above) by asking for pointers to presence services (i.e., PTR records that match "_presence._tcp.local."), querying each service for its service instance and port (i.e., SRV record), mapping each service instance to an IP address (i.e., A record), and finding out additional information about the entity using the service (i.e., TXT record parameters). As a result, Romeo's client will discover any number of local presence services, among them a service named "juliet@pronto" (with some intriguing TXT record parameters) at IP address 10.2.1.187 and port 5562. Being a romantic fellow, he then initiates a chat with you by opening an XML stream to the advertised IP address and port.

Your client then responds with a response stream header (perhaps subject to user approval -- it's not always safe to chat with strangers!).

Romeo then sends you an XMPP message.

And you reply.

You chat with Romeo for a while, then your client closes the stream.

And Romeo's client does the same.

Finally you decide to head home, so your mDNS daemon sends a Multicast DNS "Goodbye" packet for your PTR record. As a result, everyone else at the hotspot receives a Multicast DNS "Remove" event, at which point they cancel any outstanding A, SRV, TXT, or NULL record queries related to your presence service.

DNS-SD enables service definitions to include a TXT record that specifies parameters to be used in the context of the relevant service type. The name of the TXT record is the same as that of the SRV record (i.e., "username@machine-name._presence._tcp.local."). The value of the TXT record is a binary object that contains one or more strings, where (1) each string is a parameter that usually takes the form of a key-value pair and (2) the parameters are separated by a single-length byte ("0x##") that specifies the length of the parameter itself.

For detailed information about the format of the TXT record value, refer to the DNS-SD specification. The following truncated example illustrates the format.

Note: In accordance with Section 6.7 of draft-cheshire-dnsext-dns-sd, the first parameter in the TXT record value SHOULD be "txtvers".

The ®ISTRAR; maintains a registry of the parameters that can be used in the TXT record value for the _presence._tcp service type, as specified in the XMPP Registrar Considerations section of this document. Those parameters are not listed here.

It is OPTIONAL to include any of these TXT record parameters, and an implementation MUST NOT fail (i.e., MUST enable serverless messaging) even if none of the parameters are provided by another entity.

Most of the registered TXT record parameters relate to human users, in which context certain parameters are of greater interest than others, e.g. "msg", "nick", and "status"; however, serverless messaging can be used by non-human entities (e.g., devices).

Note: See the Security Considerations section of this document regarding the inclusion of information that can have an impact on personal privacy (e.g., the "1st", "last", "nick", "email", and "jid" parameters).

Devices that use serverless messaging can have multiple network interfaces. As a result, it is possible to discover the same entity multiple times. Even if a client discovers the same presence name on multiple network interfaces, it MUST show only one entity in the serverless roster. In addition, because local IP addresses can be dynamically re-assigned, the client SHOULD NOT store the IP address to be used for communication when it discovers that address in the initial DNS lookup phase; instead, it SHOULD delay sending the Multicast DNS query until the client is ready to communicate with the other entity.

If an entity has an associated icon (e.g., a user avatar or photo), its client SHOULD publish the raw binary data for that image via a DNS NULL record of the following form:

Note: In accordance with &rfc1035;, the data MUST be 65535 octets or less.

After retrieving the "phsh" value from a Buddy's TXT record, a client SHOULD search its local picture database to learn the last recorded picture hash value for an entity and then compare it to the "phsh" value in the TXT record. If the values are equal, the client SHOULD use the local copy of the icon. If the picture hash values are not equal, the client SHOULD issue a Multicast DNS NULL record query to retrieve the new icon. After retrieving the NULL record, the client SHOULD replace the old "phsh" value in the picture database with the new "phsh" value and save the icon to disk. If the client needs to send a Multicast DNS query in order to retrieve the icon, it MUST cancel the NULL record query immediately after receiving a response containing the new picture data.

If a user changes their picture, the user's client MUST update the NULL record with the contents of the new picture, calculate a new picture hash, and then update the "phsh" value in the TXT record with the new hash value. Since all users "logged into" serverless presence are monitoring for TXT record changes, they will see that the "phsh" value was changed; if they wish to view the new icon, their clients SHOULD issue a new Multicast DNS query to retrieve the updated picture.

The port used for serverless messaging MAY be any unassigned port number, as determined by the messaging application on the device. The chosen port MUST be specified in the SRV record and applications MUST use the port specified in the SRV record. However, the chosen port SHOULD also be specifed in the "port.p2pj" TXT record for backwards-compatibility with older implementations, and if included the port specified in the TXT record MUST be the same as the port specified in the SRV record.

Serverless messaging via the _presence._tcp DNS SRV service type is not limited to local networks, since it is possible to advertise this service type via Wide-Area DNS-SD as described at <http://www.dns-sd.org/iChatWideArea.html>. Although the protocol is most commonly used on local networks, there is nothing intrinsic to the protocol that limits its use to peers on the same link, and it also works between any two peers that can discover each other via any profile of DNS-SD (whether local or wide-area). Naturally, the DNS records used in Wide-Area DNS-SD will not contain the ".local." domain, since the records are not intended for use over a local link.

XMPP networks use TLS (&rfc4346;) for channel encryption, SASL (&rfc4422;) for authentication, and the Domain Name System (&rfc1034;) for weak validation of server hostnames; these technologies help to ensure the identity of sending entities and to encrypt XML streams. By contrast, zero-configuration networking uses dynamic discovery and asserted machine names as the basis of sender identity. Therefore, serverless messaging does not result in authenticated identities in the same way that XMPP itself does, nor does it provide for an encrypted channel between entities.

There are two potential solutions to this problem:

1. Negotiate the use of TLS and SASL for the XML stream as described in RFC 3920.
2. Negotiate encryption with identity checking for the message exchange using &xep0116;.

It is RECOMMENDED to use one of these methods to secure communications between serverless entities. However, subject to client configuration and local service policies, two entities MAY accept an unauthenticated and unencrypted channel; but a client SHOULD warn the human user that the channel is unauthenticated and unencrypted.

Because of fundamental differences between a true XMPP network and a serverless client "mesh", entities communicating via serverless messaging MUST NOT attempt to inject serverless traffic onto an XMPP network and an XMPP server MUST reject communications until an entity is properly authenticated in accordance with the rules defined in RFC 3920. However, a client on a serverless mesh MAY forward traffic to an XMPP network after having properly authenticated on such a network (e.g., to forward a message received on a serverless client mesh to a contact on an XMPP network).

Because there is no mechanism for validating the information that is published in DNS TXT records, it is possible for clients to "poison" this information (e.g., by publishing email addresses or Jabber IDs that are controlled by or associated with other users).

The TXT record parameters optionally advertised as part of this protocol MAY result in exposure of privacy-sensitive information about a human user (such as full name, email address, and Jabber ID). A client MUST allow a user to disable publication of this personal information (e.g., via client configuration).

The ®ISTRAR; maintains a registry of parameter strings contained in the TXT record advertised for serverless messaging (see &LINKLOCAL;).