ManagePrompt

Debugging an LLM-backed feature is awkward because the interesting part — the exact prompt that was sent and the raw response that came back — is usually invisible. ManagePrompt closes that gap with a small local capture layer. You add a line or two to your application, every completion call gets recorded, and a local web UI shows the request and response bodies, token counts, estimated cost and latency as the calls happen.

Integration is deliberately minimal. TypeScript and JavaScript projects can wrap their model with an AI SDK middleware, while code using the OpenAI SDK can pass requests through a capture() wrapper. Under the hood it works with any provider — OpenAI, Anthropic, Google, Mistral and others — and captures streaming and non-streaming calls alike, so you do not lose visibility on the calls that are hardest to inspect.

Everything runs locally and is stored in a project-scoped SQLite database, which means no hosted account and no prompts leaving your machine. The server component, which serves the web UI, is distributed through Homebrew and Go, while the capture package ships on npm. The project is MIT licensed and developed in the open.