New projects should use CQL. Sending CQL over the wire is not hard, but clients here provide connection pooling, failover, and often other benefits like tracing support.
For older Thrift clients, see ClientOptionsThrift.
The StorageProxy API is available to JVM-based clients, but unless you really know that you need it you should probably be using a higher-level client listed above or, failing that, the Thrift API. The StorageProxy API is intended for internal use, and highly specialized use-cases.
Running Hadoop map/reduce jobs in Cassandra is described in HadoopSupport.