Vessel

Vessel provides a local-first observability proxy for large language model (LLM) traffic, enabling comprehensive capture, search, replay, and comparison of every prompt. Key features include: • Traffic capture across Ollama, OpenAI, and Anthropic backends • Full-text search and filtering for prompts and responses • Request replay across different models and parameters • Response scoring and comparative analysis • Built-in MCP server for real-time trace queries This single binary solution runs entirely on your local machine, requiring no account and ensuring data privacy as nothing leaves your system. It meticulously captures prompts, responses, timing, and token counts, even for streamed interactions. Users can easily identify problematic requests by searching across various filters like model, backend, or agent tag, with errors and truncations clearly badged. Vessel is ideal for developers and teams working with language models who need deep insights into model behavior and performance. It allows for tagging agents and naming runs, providing granular filtering and reporting capabilities. The ability to re-send captured requests to multiple models simultaneously and compare their outputs side-by-side with performance metrics like timing and token deltas streamlines model evaluation and optimization. Furthermore, the embedded read-only MCP server empowers coding assistants to directly query traffic traces, simplifying troubleshooting and performance diagnosis.