Aims. We study the spatial clustering of 632 (1130) XMM-COSMOS active galactic nuclei (AGNs) with known spectroscopic or photometric redshifts in the range z = [0.1–2.5] in order to measure the AGN bias and estimate the typical mass of the hosting dark matter (DM) halo as a function of AGN host galaxy properties. Methods. We created AGN subsamples in terms of stellar mass, M_*, and specific black hole accretion rate, L_X/M_*, to study how AGN environment depends on these quantities. Further, we derived the M_*−M_halo relation for our sample of XMM-COSMOS AGNs and compared it to results in literature for normal non-active galaxies. We measured the projected two-point correlation function w_p(r_p) using both the classic and the generalized clustering estimator, based on photometric redshifts, as probability distribution functions in addition to any available spectroscopic redshifts. We measured the large-scale (r_p ≳ 1 h⁻¹ Mpc) linear bias b by comparing the clustering signal to that expected of the underlying DM distribution. The bias was then related to the typical mass of the hosting halo M_halo of our AGN subsamples. Since M_* and L_X/M_* are correlated, we matched the distribution in terms of one quantity and we split the distribution in the other. Results. For the full spectroscopic AGN sample, we measured a typical DM halo mass of log (M_halo/h⁻¹ M_⊙) = 12.79_−0.43^+0.26 , similar to galaxy group environments and in line with previous studies for moderate-luminosity X-ray selected AGN. We find no significant dependence on specific accretion rate L_X/M_*, with log (M_halo/h⁻¹ M_⊙) = 13.06_−0.38^+0.23 and log (M_halo/h⁻¹ M_⊙) = 12.97_−1.26^+0.39 for low and high L_X/M_* subsamples, respectively. We also find no difference in the hosting halos in terms of M_* with log (M_halo/h⁻¹ M_⊙) = 12.93_−0.62^+0.31 (low) and log (M_halo/h⁻¹ M_⊙) = 12.90_−0.62^+0.30 (high). By comparing the M_*−M_halo relation derived for XMM-COSMOS AGN subsamples with what is expected for normal non-active galaxies by abundance matching and clustering results, we find that the typical DM halo mass of our high M_* AGN subsample is similar to that of non-active galaxies. However, AGNs in our low M_* subsample are found in more massive halos than non-active galaxies. By excluding AGNs in galaxy groups from the clustering analysis, we find evidence that the result for low M_* may be due to larger fraction of AGNs as satellites in massive halos.