Background Transcutaneous auricular vagus nerve stimulation (taVNS) is a promising noninvasive intervention for mild cognitive impairment (MCI), but its clinical efficacy and network-level mechanisms remain unclear. Methods In this prospective, randomized, sham-controlled, parallel-group trial, 60 patients with MCI were assigned 1:1 to 12-week taVNS or sham taVNS. The primary outcome was the between-group difference in change from baseline to week 12 on the Montreal Cognitive Assessment-Basic (MoCA-B). Secondary outcomes included AVLT-H, STT-A/B, AFT, BNT, HAMD-17, and HAMA scores. Twenty-eight healthy controls underwent baseline rs-fMRI as a non-randomized reference cohort. Functional connectivity (FC) identified MCI-related abnormal regions, which together with theory-driven regions served as seeds for Granger causality analysis (GCA) to characterize intervention-related effective connectivity changes. Imaging-behavior associations were examined using Pearson correlations. The final per-protocol analysis included 52 randomized patients with complete clinical and imaging data. Results After 12 weeks, the taVNS group showed significant within-group improvements in global cognition and multiple secondary outcomes, whereas the sham group did not. The between-group difference in MoCA-B change was significant (mean difference = 2.91, 95% CI: 0.79 to 5.03, Cohen’s d = 0.59), indicating a moderate-to-large effect. Baseline FC abnormalities in MCI mainly involved the insula, medial frontal cortex, precuneus, temporal cortex, thalamus, cingulate cortex, and parieto-occipital regions. Following taVNS, effective connectivity was selectively reconfigured within a distributed network centered on the insula, anterior cingulate cortex, temporal cortex, and precuneus. Changes in directed connectivity correlated with changes in MoCA-B, AFT, and HAMA scores. No serious adverse events occurred. Clinical trial registration https://www.chictr.org.cn/showproj.html?proj=131823 , identifier (ChiCTR2100049940).