Stanford University's X-lab and SNAP Lab have jointly released ProteoCraft, a protein design framework that introduces a "propose-verify" search scheme: approximate gradients guide the direction of sequence modifications, full structure prediction verifies candidates, and a two-stage Monte Carlo tree search orchestrates broad multi-position exploration followed by single-point local refinement—supporting both de novo protein design and targeted local modification of defined regions. Multi-round computational benchmarks show that its standalone AF2-IG evaluation achieves a strict pass rate of 86.2%; under the same candidate budget, the tree search's high-confidence pass rate is 2.85 times that of the Best-of-N control; and the optimized Big configuration delivers an average 3.57-fold end-to-end search speedup with essentially unchanged design quality. In wet-lab experiments, of 35 successfully expressed candidates for PD-L1 mini-protein design, 29 were binding hits—a hit rate of 82.9%—with the best candidate showing an SPR-fitted Kd of 0.879 nM; the hit candidates exhibited sufficient sequence diversity and performed well in assessments related to immunogenicity and off-target risk. The team also obtained validation results in mini-protein and VHH nanobody design against the IFNAR2 target, as well as in local CDR modification tasks. The team is currently exploring function-first, on-demand development of immunomodulatory proteins based on this framework.