Building computational models of complex systems, including ones governing the behavior of biological cells, is a laborious process involving manual information gathering and model implementation. This makes it difficult to build models based on known causal mechanisms, creating a gap between the scope of typical models, and the scale of data they are meant to help interpret. We developed INDRA (the Integrated Network and Dynamical Reasoning Assembler), a system for automatically assembling mechanistic models directly from English language, including from the scientific literature. INDRA interfaces with natural language processing systems to extract descriptions of mechanisms from text, and uses knowledge assembly algorithms to fix certain errors, resolve redundant knowledge, infer missing information, and assess belief. INDRA then uses this knowledge base to produce models in different formalisms, including rule-based dynamical models, and causal graphs. We present applications of INDRA including (i) a model automatically built from an English language description of mechanisms which can reproduce resistance to a melanoma cancer drug (ii) a model of molecular signaling pathways built by machine-reading ~95,000 scientific publications, which can provide mechanistic explanations to drug perturbations in a cancer cell line (iii) a human-machine dialogue system in which a user can gather information and build a mechanistic hypothesis for an observed phenomenon by talking with a computer partner.