llmpress reduces the token cost of sending source code to an LLM by detecting repeated patterns, replacing them with short aliases, and transparently restoring the originals in the model's response.