Background: Capecitabine (cape), an oral chemotherapy, is the treatment backbone for many GI cancers. Its complex dosing and narrow therapeutic index make medication adherence and toxicity management crucial for quality patient care. Methods: We conducted a feasibility study of “Penny,” an augmented intelligence mobile phone chatbot that leverages algorithmic surveys and natural language processing (NLP) to engage with patients in conversational, bi-directional text messages. Penny provides patients with medication reminders tailored to their prescribed doses and schedules, sends weekly check-in messages, manages low-grade symptoms in real time, and escalates high-grade symptoms for resolution by the clinical team. Patients ≥18 years old receiving cape for the treatment of a GI cancer were accrued in sequential cohorts of 20 for participation over a three-month period. Feasibility was assessed during planned interim analyses and was predefined as the completion of a 20-patient cohort without a safety event, defined as the communication of incorrect medication or symptom management recommendations as ascertained by two independent clinician reviewers (Κ = 0.89). Secondary outcomes included patient-reported adherence and engagement with the chatbot’s weekly check-in messages. At study completion, all patients were invited to participate in structured interviews to provide feedback on the platform. Results: The first cohort of 20 patients was enrolled from 8/2021 to 4/2022; the median age was 57 years, and patients were primarily female (55%), white (65%), commercially insured (55%), and had colorectal cancer (55%). Chemotherapy regimens included cape with oxaliplatin (50%), concurrent RT (30%), temozolomide (5%), and monotherapy (15%). A total of 2,149 text messaging exchanges were reviewed with 150 (7%) medication-related and 9 (0.4%) symptom-related safety events identified. Most medication-related safety events were due to misalignment with prescribed chemotherapy schedules (55%) and doses (32%). Symptom-related safety events were primarily due to the misinterpretation of patient messages by Penny’s NLP functionality (89%). Average patient-reported adherence was 67% (SD 27%), and patients engaged with 27% (SD 24%) of the chatbot’s weekly check-in messages. In post-study interviews with 12 patients, participants reported that the medication reminders were reliable and user-friendly, whereas the symptom management tool was too simplistic to be helpful. Conclusions: Although Penny has not yet met its feasibility endpoint, the lessons learned from this first cohort have informed further refinements to the platform. Ongoing efforts aim to integrate Penny with the electronic health record and further train the chatbot’s NLP functionality to minimize medication- and symptom-related safety events, respectively. Clinical trial information: NCT05113264.