The therapeutic mechanisms of brain-computer interface (BCI) rehabilitation remain partially understood. Specifically, it is unclear whether functional gains are driven primarily by active neural engagement or strictly by the temporal contingency between motor intention and sensory feedback. This study aimed to deconstruct the BCI paradigm to isolate the specific contributions of these core components to functional recovery and cortical reorganization in subacute stroke. Seventy-seven subacute stroke patients were randomized into three intervention arms. Group C (synchronous) received robotic assistance strictly contingent on real-time detection of task engagement during motor imagery; Group B (asynchronous) performed identical motor imagery tasks but received temporally uncoupled robotic assistance; and Group A (sham) received sham motor imagery detection and preset passive robotic assistance. The intervention lasted for 3 weeks. The primary clinical outcome was Fugl-Meyer Assessment-Upper Extremity (FMA-UE) scores. Functional near-infrared spectroscopy (fNIRS) was included as an exploratory neuroimaging measure to provide preliminary information on cortical responses associated with the intervention. Sixty-eight participants were included in the final analysis. After adjustment for baseline values, Group C showed significantly higher post-intervention FMA-UE scores than both Group A (adjusted mean difference = 7.47, 95% CI [1.94, 13.00], P = 0.004) and Group B (adjusted mean difference = 9.72, 95% CI [4.14, 15.30], P < 0.001). Group C also showed significantly higher post-intervention MBI scores than both Group A (adjusted mean difference = 21.15, 95% CI [12.49, 29.81], P < 0.001) and Group B (adjusted mean difference = 21.27, 95% CI [12.42, 30.12], P < 0.001). Exploratory neuroimaging analyses showed that only the change in ipsilesional sensorimotor cortex laterality index during motor imagery was significantly correlated with motor improvement ( ρ = 0.404, P = 0.002, q = 0.012), whereas no resting-state or task-related fNIRS group comparisons remained significant after false discovery rate correction. Synchronous BCI training was associated with superior clinical improvement compared with asynchronous and sham conditions in subacute stroke. The benefit was observed only when robotic assistance was contingent on real-time task performance, despite all three groups receiving comparable device exposure and motor imagery instruction. The neuroimaging results remain exploratory and should be confirmed in larger studies with prospectively defined imaging endpoints. Trial registration This trial was registered under the Chinese Clinical Trial Registry (ChiCTR2400091197) and was prospectively registered on October 21, 2024. Ethics Approval: Approved by the Ethics Committee of Xuzhou Rehabilitation Hospital (XK-LSW-20241017-009).