Globally, severe acute malnutrition (SAM), defined as a weight-for-length z-score more than three SDs below a reference mean (WLZ < -3), affects 14 million children under 5 years of age. Complete anthropometric recovery after standard, short-term interventions is rare, with children often left with moderate acute malnutrition (MAM; WLZ -2 to -3). We conducted a randomized controlled trial (RCT) involving 12- to 18-month-old Bangladeshi children from urban and rural sites, who, after initial hospital-based treatment for SAM, received a 3-month intervention with a microbiome-directed complementary food (MDCF-2) or a calorically more dense, standard ready-to-use supplementary food (RUSF). The rate of WLZ improvement was significantly greater in MDCF-2-treated children (P = 8.73 × 10-3), similar to our previous RCT of Bangladeshi children with MAM without antecedent SAM (P = 0.032). A correlated meta-analysis of plasma levels of 4520 proteins in both RCTs revealed 215 positively associated with WLZ (largely representing musculoskeletal and central nervous system development) and 44 negatively associated (primarily related to immune activation). Moreover, the positively associated proteins were significantly enriched by MDCF-2 (q = 1.1 × 10-6). Characterizing the abundances of 754 bacterial metagenome-assembled genomes in serially collected fecal samples disclosed the effects of acute rehabilitation for SAM on the microbiome and how, during treatment for MAM, specific strains of Prevotella copri function at the intersection between MDCF-2 glycan metabolism and anthropometric recovery. These results provide a rationale for further testing the generalizability of MDCF efficacy and for identifying biomarkers to define treatment responses.