Abstract: In low-carbon agri-food supply chains, centralized optimization is impractical since entities are often reluctant to disclose internal parameters such as costs and loss rates due to commercial confidentiality. To address this issue, a distributed configuration optimization method based on responsibility tokens and the federated alternating direction method of multipliers (FedADMM) is proposed. Two token transfer payments, namely time tokens and demand tokens, are introduced in a configuration model with carbon cap-and-trade constraints. Three categories of cross-entity couplings—service time, cumulative cost, and demand with loss—are thereby relaxed into economic penalties for contract violations. Token prices are then proved equivalent to the ADMM dual multipliers. Based on this, a FedADMM algorithm is built in which only aggregated interface variables are exchanged during each iteration, thereby preserving the privacy of core parameters such as costs. An apple supply chain is used as a case study. The proposed FedADMM obtains the same optimal solution as the centralized method, while improving computational efficiency by about an order of magnitude and reducing the parameter exposure rate from 100% to zero. Sensitivity analysis further shows that the elasticity of total cost with respect to the loss rate is about two orders of magnitude greater than that with respect to carbon-policy parameters. The optimal configuration remains unchanged under simultaneous ±50% perturbations in demand and carbon price. The proposed method reaches the global optimum while preserving the privacy of participating entities, offering an effective way for distributed configuration optimization in low-carbon agri-food supply chains.