We introduce MutEval, a reproducible dual-mode framework for evaluating the robustness of code generation mod- els. MutEval examines: (i) prompt-level robustness by applying semantically equivalent paraphrases and syntactic perturbations to natural-language prompts, and (ii) code-level robustness by introducing small, controlled mutations to task descriptions or partial code contexts. Built on the HumanEval benchmark, the framework generates deterministic mutations, queries LLMs, and measures robustness using pass@k, CodeBERTScore, and similarity metrics. Across four state-of-the-art code LLMs, we ob- serve consistent degradation patterns. Mutated natural-language prompts lead to moderate drops in similarity (63–68%) and pass@1 (8-18%), with gpt-4o variants showing the strongest resilience. Mutated code contexts, however, are substantially more harmful: similarity falls to 7–15% and Pass@1 declines to 21–32%, over 50% below original performance, despite CodeBERTScore remaining above 79%. These results highlight that code models are significantly more brittle to perturbations in code than in natural language and emphasize the impor- tance of robustness-aware evaluation for reliable deployment. A video demonstration of MutEval is available at https://youtu.be/v5mc8wneqRA.