Most results from humann3 unclassified

I am trying to run humann3 (v. 3.7) on acollection of 537 good quality (>50% complete, < 5%redundant) MAGs assembled from human gut samples. The MAGs were assembled by metabat3. However, the results look very odd. 84% of the MAGs (451) return completely empty results:

Pathway 171007M.6_Abundance
UNMAPPED 0.0000000000
UNINTEGRATED 0.0000000000

and the few that return some results at all (86) have very few lines:

Pathway 171003M.19_Abundance
UNMAPPED 0.0000000000
UNINTEGRATED 38.2009166740
UNINTEGRATED|unclassified 38.2009166740
PWY-7238: sucrose biosynthesis II 0.7363770250
PWY-7238: sucrose biosynthesis II|unclassified 0.7363770250

The command I am running is for each MAG file (fasta):
humann --input MAG_name.fa --output OUT_DIR

The slurm output file has a lot of warnings of this sort:

Total species selected from prescreen: 0

Selected species explain 0.00% of predicted community composition

No species were selected from the prescreen.
Because of this the custom ChocoPhlAn database is empty.
This will result in zero species-specific gene families and pathways.

I believe I have downloaded and am using the full chocophlan and uniref databases.

What am I doing wrong here? Does humann have to be run at the read level?

HUMAnN is designed for profiling molecular functions from unassembled shotgun metagenomes (i.e. with your sequencing reads as input). It sounds like you’re trying to do something more like genome annotation of your MAGs? I would recommend an annotation workflow such as Prokka for that task.