Your coding agent can now watch itself think and write its own exams

What it does

Raindrop Workshop is a local trace viewer for AI coding agents. It mirrors every token, tool call, and span into a web UI as it happens, giving you a live window into your agent’s decision-making. The captured traces are also readable by the agent itself, allowing it to author assertions against your codebase and iterate until they pass.

The interesting bit

The project treats eval-writing as part of the agent’s job, not the developer’s. A “self-healing” loop lets the agent author an eval, run itself, watch it fail, patch the code, and rerun until assertions pass—essentially making the agent responsible for its own regressions.

Key highlights

• Live-streamed traces with no polling or refresh; data appears the moment the agent emits it
• One-command local install that requires no cloning, building, or dependency wrangling
• Broad compatibility across languages (TypeScript, Python, Go, Rust), SDKs (Vercel AI, OpenAI, Anthropic, LangChain, etc.), and coding agents (Claude Code, Codex, Cursor, etc.)
• Optional local replay endpoint that can replay a production trace against real agent code
• Companion cloud product (Raindrop Cloud) for production observability, using the same binary

Verdict

Worth a look if you’re running agentic code locally and tired of black-box reasoning. Skip it if you don’t use a supported coding agent or if your observability needs are already met by your existing tracing stack.