Moorlaas (1928) proposed that apraxic patients can identify objects and can remember the purpose they have been made for but do not know the way in which they must be used to achieve that purpose. Knowledge about the use of objects and tools can have two sources: It can be based on retrieval of instructions of use from semantic memory or on a direct inference of function from structure. The ability to infer function from structure enables subjects to use unfamiliar tools and to detect alternative uses of familiar tools. It is the basis of mechanical problem solving. The purpose of the present study was to analyze retrieval of instruction of use, mechanical problem solving, and actual tool use in patients with apraxia due to circumscribed lesions of the left hemisphere. For assessing mechanical problem solving we developed a test of selection and application of novel tools. Access to instruction of use was tested by pantomime of tool use. Actual tool use was examined for the same familiar tools. Forty two patients with left brain damage (LBD) and aphasia, 22 patients with right brain damage (RBD) and 22 controls were examined. Only LBD patients differed from controls on all tests. RBD patients had difficulties with the use but not with the selection of novel tools. In LBD patients there was a significant correlation between pantomime of tool use and novel tool selection but there were single cases who scored in the defective range on one of these tests and normally on the other. Analysis of LBD patients' lesions suggested that frontal lobe damage does not disturb novel tool selection. Only LBD patients who failed on pantomime of object use and on novel tool selection committed errors in actual use of familiar tools. The finding that mechanical problem solving is invariably defective in apraxic patients who commit errors with familiar tools is in good accord with clinical observations, as the gravity of their errors goes beyond what one would expect as a mere sequel of loss of access to instruction of use.