How about avro-tools-VERSION-nodeps.jar, with a maven classifier of 'nodeps'? That should be pretty easy to do with a minor pom change, I'll have a look at it in a few days.
The thrift-protobuf-compare benchmark requires it, and others call the tools classes directly instead of by command-line.
Off-topic FYI RE benchmarks:
There are some changes to the benchmark code and Avro that can speed up the thrift-protobuf-compare stuff. If Utf8 is 'more lazy' and doesn't create a byte array initially when constructed from a String, that helps. Replacing some of the
for (type thing : things) loops in the harness code with
for (int counter; condition; increment) loops helps some. The harness overhead is high, time is mostly in String to Utf8 conversions, GenericDatumReader/Writer, and the harness – almost nothing is in the low level serialization/deserialization. Because of the requirement to extract their object from our IndexedRecords, the Reflect API is probably fastest here – we can avoid the extra copy step and produce the exact object type.