OBJECTIVE: To determine the short-term and long-term risks of tuberculosis (TB) associated with CD4 cell recovery during antiretroviral therapy (ART). DESIGN: Observational community-based ART cohort in South Africa. METHODS: TB incidence was determined among patients (n = 1480) receiving ART for up to 4.5 years in a South African community-based service. Updated CD4 cell counts were measured 4-monthly. Person-time accrued within a range of CD4 cell count strata (CD4 cell strata) was calculated and used to derive CD4 cell-stratified TB rates. Factors associated with incident TB were identified using Poisson regression models. RESULTS: Two hundred and three cases of TB were diagnosed during 2785 person-years of observation (overall incidence, 7.3 cases/100 person-years). During person-time accrued within CD4 cell strata 0-100, 101-200, 201-300, 301-400, 401-500 and more than 500 cells/microl unadjusted TB incidence rates were 16.8, 9.3, 5.5, 4.6, 4.2 and 1.5 cases/100 person-years, respectively (P < 0.001). During early ART (first 4 months), adjusted TB rates among those with CD4 cell counts 0-200 cells/microl were 1.7-fold higher than during long-term ART (P = 0.026). Updated CD4 cell counts were the only patient characteristic independently associated with long-term TB risk. CONCLUSION: Updated CD4 cell counts were the dominant predictor of TB risk during ART in this low-resource setting. Among those with baseline CD4 cell counts less than 200 cells/microl, the excess adjusted risk of TB during early ART was consistent with 'unmasking' of disease missed at baseline screening. TB incidence rates at CD4 cell counts of 200-500 cells/microl remained high and adjunctive interventions are required. TB prevention would be improved by ART policies that minimized the time patients spend with CD4 cell counts below a threshold of 500 cells/microl.