Triangle counting is a fundamental problem in the analysis of large graphs. There is a rich body of work on this problem, in varying streaming and distributed models, yet all these algorithms require reading the whole input graph. In many scenarios, we do not have access to the whole graph, and can only sample a small portion of the graph (typically through crawling). In such a setting, how can we accurately estimate the triangle count of the graph? We formally study triangle counting in the {\em random walk} access model introduced by Dasgupta et al (WWW '14) and Chierichetti et al (WWW '16). We have access to an arbitrary seed vertex of the graph, and can only perform random walks. This model is restrictive in access and captures the challenges of collecting real-world graphs. Even sampling a uniform random vertex is a hard task in this model. Despite these challenges, we design a provable and practical algorithm, TETRIS, for triangle counting in this model. TETRIS is the first provably sublinear algorithm (for most natural parameter settings) that approximates the triangle count in the random walk model, for graphs with low mixing time. Our result builds on recent advances in the theory of sublinear algorithms. The final sample built by TETRIS is a careful mix of random walks and degree-biased sampling of neighborhoods. Empirically, TETRIS accurately counts triangles on a variety of large graphs, getting estimates within 5\% relative error by looking at 3\% of the number of edges.