That’s going to be quite challenging… Other people with low b-value single-shell data have reported some success by adding a CSF compartment; but I would highly advise to leave out the GM one. See this post (link) for a long description of options I gave to a few other users, and some results posted by a user further on in that thread (scroll up to get more context, scroll down for results).
In the interest of getting a more accurate CSF response function, I would advise to use the dhollander algorithm in dwi2response as well; we’ve often found that (apart from being more convenient to use; no 5tt segmentation and registration of T1w image to diffusion data required) specifically the CSF response function from the msmt_5tt algorithm can be (quite) suboptimal. The dhollander algorithm doesn’t seem to have this problem.
Using the instructions in the aforementioned post, you could go for dwi2response dhollander followed by dwi2fod msmt_csd, but the latter one with only the WM and CSF responses (ditching the GM one at this stage); and then followed by tckgen for probabilistic streamline tractography (which will allow you to use the ACT framework for anatomical constraints as well if you provide it with your 5tt segmentation; this should particularly benefit you well at a low b-value and in the absence of a GM compartment).
If you want to go for tckglobal, it may be worth a shot trying this as well without the GM response (so again just the WM and CSF one)… but I’m not sure how well that would work for such limited data (low b-value and limited number or gradient directions). @dchristiaens can probably comment further on this option.