Introduction Higher physical activity (PA) is associated with lower risk of several cancers, but the biological mechanisms underlying these associations remain unclear. We integrated observational and Mendelian randomization (MR) approaches to investigate circulating proteins as potential intermediates linking PA to cancer risk. Methods We analysed accelerometer-derived PA measures (~100,000 participants), circulating protein levels (~53,000 participants) derived using the Olink Explore 3072 panel and cancer incidence (breast, colorectal, endometrial, and prostate cancers) via linked registry data among UK Biobank (UKB) participants. Observational associations were assessed using regression models and Cox proportional hazard models adjusting for demographic, lifestyle, and clinical covariates, including body mass index (BMI). Proteins associated with both overall acceleration average and cancer incidence were taken forward in a two-sample and two-step MR analyses using genome-wide association data. Results Higher overall acceleration average was observationally associated with lower risk of breast, colorectal, and endometrial cancer, with attenuation after BMI adjustment for endometrial and colorectal cancer, and a positive association with prostate cancer. MR analyses supported a protective causal effect of higher overall PA on breast, colorectal, and prostate cancer risk, and suggested a much larger effect than observational analyses. Overall, PA was observationally associated with 436 circulating proteins of which 44 were also associated with incident cancer risk. MR analyses provided some support that TNFSRF13B was influenced by physical activity in the same direction as the observational analyses but found conflicting evidence for an effect of TNFSRF13B on breast cancer, although the MR analysis is consistent with a protective effect of physical activity on breast cancer risk. Conclusions While higher overall physical activity appears to causally reduce risk of several cancers, circulating proteins associated with PA largely do not show strong evidence of being intermediates of these effects, suggesting they may act primarily as biomarkers of physical activity rather than causal intermediates.