The geometry of metal-ligand interactions in proteins is examined and compared with information for small-molecule complexes from the Cambridge Structural Database (CSD). The paper deals with the metals Ca, Mg, Mn, Fe, Cu, Zn and with metal-donor atom distances, coordination numbers and extent of distortion from ideal geometry (octahedral, tetrahedral etc.). It assesses the agreement between geometry found in all metalloprotein structures in the Protein Data Bank (PDB) determined at resolution < or = 1.6 A with that predicted from the CSD for ligands which are analogues of amino-acid side chains in proteins [Harding (1999), Acta Cryst. D55, 1432-1443; Harding (2000), Acta Cryst. D56, 857-867]. The agreement is reasonably good for these structures but poorer for many determined at lower resolution (examined to 2.8 A resolution). For metal-donor distances, the predictions from the CSD, with minor adjustments, provide good targets either for validation or for restraints in refinement of structures where only poorer resolution data is available. These target distances are tabulated and the use of restraints is recommended. Validation of angles or the use in refinement of restraints on angles at the metal atom is more difficult because of the inherent flexibility of these angles. A much simplified set of parameters for angle restraints with quite large standard deviations is provided. (Despite the flexibility of the angles, acceptable and preferred coordination numbers and shapes are well established and a summary table is provided.) An unusual and perhaps biochemically important feature of Zn coordination with carboxylate seen in the CSD examples is also clearly present in metalloprotein structures. With metals like Ca, carboxylate coordination is monodentate or bidentate (two M-O bonds of nearly equal length). In Zn carboxylates a continuous range between monodentate and bidentate coordination is found, with one Zn-O bond of normal length and another of any length between this and a van der Waals contact.