Long Range Wide Area Network (LoRaWAN) covers the needs of energy-constrained IoT-devices for operational longevity and extended communication range in a best-effort fashion. However, Lo- RaWAN’s minimalist design cannot handle the traffic from dense deployments with more than a few hundred devices connected to a single gateway, since each LoRa-device transmits data-packets without any information regarding the availability of the medium. In this paper, we try to improve the scalability of LoRaWAN by manifolds, serving thousands of devices per gateway. We present a novel protocol called p persistent-Channel Activity Recognition Multiple Access (p-CARMA) that exploits LoRaWAN’s Channel Activity Detection (CAD) as a crude mechanism to assess if the channel is free. Due to CAD’s imperfections (it only scans for preambles, not for any channel activity) p-CARMA operates probabilistically with each device deciding on a p value based upon local estimation. At the beginning of operation, this estimate is derived from pure local information, that is without involvement of the gateway, and devices automatically adapt to changes in the environment. Then, the adaptation of p-value is assisted by critical information on the cumulative device-delays, multicasted by the gateway at regular, large timespans. To evaluate the performance of p-CARMA, we implemented it in ns-3 based upon a detailed characterization of LoRaWAN’s CAD mechanism involving an extensive set of real-world experiments. We compared p-CARMA to vanilla LoRaWAN as well as a variant using the theoretically optimal p = 1=N (N being the total number of devices). The simulation results show that p-CARMA achieves from three-fold, up to a twenty-fold higher Packet Reception Ratio than LoRaWAN while handling thousands of devices. Further, its adaptivity outperforms the fixed p-value by a factor of 5.25 when scaling up. Moreover, p-CARMA does so while consuming 37.31%-58.17% less energy on average per device compared to vanilla LoRaWAN. @en