Calleja-Rodelas et al. present a unique dataset of land-atmosphere fluxes of energy, water, and carbon for agroforestry systems and adjacent conventional systems. They use a novel low-cost eddy covariance technique and use state-of-the-art post-processing procedures. In my opinion, the dataset provides insights into ecosystems that are understudied given their potential for climate mitigation. The manuscript is well written and provides enough detail to understand the processing steps. My concern is that ancillary data is missing that is crucial for the interpretation of the fluxes to assess climate mitigation potential. First, without information on standing biomass, carbon export through harvest, and carbon import through organic fertiliser, the existing flux data will be difficult to interpret. The authors provide a table with information on crop rotation, but I believe this is insufficient. At the minimum, the authors should discuss how this issue can be addressed in studies using this dataset. Second, the authors mention the issue of tower location dependency of flux measurements. This is an important issue and should be addressed in more detailed. They provide yearly flux footprint and highlight the need for additional flux footprint modelling. However, to maximise the insights gained from footprint models, the authors should consider publishing spatial maps/data on land cover and properties surrounding the flux towers. End users can then use these maps to conduct their own footprint analyses.

Please find below some more specific comments:

Line 8: How do the climatic regions differ?

Table 1: Why not using the reference period 1991-2020?

Table 1: The impact from the different soil properties on fluxes could also be discussed. Mariensee stands out with a very high soil organic carbon content.

Line 89: Trees were harvested but information on exported biomass/carbon is missing.

Figure 2: The target area for Mariensee is much smaller than the flux footprint. How useful are the flux measurements then to understand the agroforestry impact? Additionally, the flux footprint overlap and fluxes are thus not independent between AF and OC/OG.

Line 183: For testing the gap-filling performance (and particularly for testing XGBoost), how was the dataset split in training and test? Was it done randomly or blocked? Due to the high autocorrelation in eddy covariance data, random selection usually results in much higher correlation and lower RMSE.

Line 289: It is important to establish if difference between lower-cost and EC methods are truly random or if a systematic bias exists. This information is important to better understand the nature of the calculated error.

Line 300: Here, and throughout the manuscript, it would be helpful to quantitatively support the main text with actual numbers. As it is written now, most statements remain qualitative.

Line 308: Why did the daytime partitioning fail? This information would be useful.

Table 6: I am not sure how meaningful the mean parameters for all variables other than zm, z, and ha are.

Table 7: I would suggest to only compare meteorological conditions between sites for overlapping periods.

Figure 3: It looks like the wind speed is decreasing over time. Is that due to the growing trees? These are important findings and should be discussed.

Line 420: I would suggest to not report partial years.