I have heavy INet restrictions at office (can't download files larger than 2Mb and big 400-500Mb traffic of multiple small files will also be punished) while I would like to participate in distributed computing. I see that many projects like climateprediction.net, rosetta etc have large files (3-20Mb+) in downloads required, and total task download traffic can consume these 400-500Mb.
How can I download task files separately and feed them to BOINC so it will start to do the task ? What about output data size estimation, will it be of same size so require separate uploading procedure again ?
Can you advise me API documentation on tasks requesting ?
I hope such out-of-BOINC download can/will be implemented.

Climateprediction has several types of model, all with different sizes of files, and usually with several files to download.

The model with the shortest runtime, also has the largest download/upload, something that was predicted years ago when people complained about the long models then available.

For this short model, the downloads consist of a basic file of about 80 Megs, plus a file of 28 Megs for each different year to be modelled, plus some smaller files.
For the upload, there are 3 files: 25 Megs, 20 Megs, and 212 K.

This project isn't suitable for people with limited bandwidth, or file size restrictions.

You can try using the following chart to choose projects that don't require much data transfer. Note: chart was last updated over 6 months ago.

http://www.boinc.cz/img/BOINC_Projects_table.png

I would nominate AQUA as a notably low-bandwidth project. The biggest hit - well above your limit - would be downloading BOINC itself, but AQUA is an order of magnitude below that. And if you download at the right time, their most demanding application can run for up to 10 days without refueling on a dual core. [Currently available tasks run shorter, and the mix changes frequently as the research develops - keep an eye on their web site].