| I’ve been building a lot with LLMs lately and kept thinking: why doesn’t this tool exist? The workflow usually ends up being: write some code, run it, tweak a prompt, add logs just to understand what actually happened. It works in some cases, breaks in others, and it’s hard to see why. You also want to know that changing a prompt or model didn’t quietly break everything. Reticle puts the whole loop in one place. You define a scenario (prompt + variables + tools), run it against different models, and see exactly what happened - prompts, responses, tool calls, results. You can then run evals against a dataset to see whether a change to the prompt or model breaks anything. There’s also a step-by-step view for agent runs so you can see why it made a decision. Everything runs locally. Prompts, API keys, and run history stay on your machine (SQLite). Stack: Tauri + React + SQLite + Axum + Deno. Still early and definitely rough around the edges. Is this roughly how people are debugging LLM workflows today, or do you do it differently? |
Here’s a quick demo of the agent execution view:
https://raw.githubusercontent.com/fwdai/reticle/main/.github...