Part one: Overview and explanation

Because tor is an application-level proxy, it needs client-side support
from every client program that wants to use it. (This is different from
systems like Freedom, which used a single client-side program to capture
all packets and redirect them to the Freedom network.) Client applications
need two general classes of modifications to be compatible with tor:

1) Whenever they call connect(), they instead should connect() to the
local onion proxy and tell it "address and port". The onion proxy will
itself make a connection to "address and port", and then the client
application can talk through that socket as if it's directly connected. To
support as many applications as possible, tor uses the common "socks"
protocol which does exactly the above. So applications with socks support
will support tor without needing any modifications.

2) Applications must not call gethostbyname() to resolve an address
they intend to later connect() to via onion routing. gethostbyname()
contacts the dns server of the target machine -- thus giving away the
fact that you intend to make an anonymous connection to it.

To clarify, I need to explain more about the socks protocol. Socks
comes in three flavors: 4, 4a, and 5. The socks4 protocol basically
uses IP and port -- so it is unsuitable because of the gethostbyname()
issue above. Socks4a is a slight modification to the socks4 protocol,
whereby you can specify an IP of 0.0.0.x to signal the socks server
that you will instead be sending a hostname (fqdn). So applications with
socks4a support are all set. Socks5, on the other hand, allows the client
to specify "address type" and then an address -- so some applications
choose to supply an IP and others choose to supply a hostname. If the
application uses socks5 you must investigate further to decide whether
it's leaking anonymity.


    Part two: using tsocks to transparently replace library calls

tsocks (available from http://tsocks.sourceforge.net/ or from your
favorite apt-get equivalent) allows you to run a program as normal,
but it replaces the system calls for connect() to connect to the socks
server first and then pass it your destination info. In our case the
socks server is a tor process (running either locally or elsewhere).
In general this works quite well for command-line processes like finger,
ssh, etc. But there are a couple of catches: A) tsocks doesn't intercept
calls to gethostbyname. So unless you specify an IP rather than hostname,
you'll be giving yourself away. B) Programs which are suid root (or
anybody else) don't let you intercept the system calls -- ssh falls into
this category. But you can make a local copy of ssh and use that. C)
Probably tsocks doesn't behave well for behemoths like Mozilla.


    Part three: applications which support tor correctly