Your AI Agent just produced a weird result. It picked a tool you didn't expect, output a semantically vague response. Logs show API calls and token counts but give no clue which decision went wrong. Raindrop launched an open-source tool on May 14 to change that: Workshop v0.1.6, a fully local, free debugger for AI Agents. It lets developers track every token output and tool call in real time, then hands the debugging itself over to Claude Code or Codex.
Why Debugging AI Agents Is Different from Traditional Software
Classical debugging relies on breakpoints, a full call stack, and deterministic execution paths. AI Agent behavior is probabilistic: the same input can lead to wildly different outcomes across runs. Decisions are scattered across multiple LLM calls, making terminal output nearly useless for root-cause analysis. The problem is not "which line of code is wrong" but "which step caused the agent to make an unexpected judgment under a specific context." Current approaches fall into two camps: cloud monitoring platforms that send traces to third-party dashboards, or custom logging infrastructure that must be maintained with every framework upgrade. Both inform developers "what happened" but don't fix it. Workshop takes a third approach: purely local execution, no data leaves the machine, open source, free, and AI actively participates in the debug loop.
How Workshop Works
Once running, Workshop launches a visual interface locally and exposes an MCP (Model Context Protocol) server. MCP is the standard protocol that lets AI coding tools like Claude Code call external capabilities. After connecting a supported SDK, every execution node of the agent—every token output, tool call, decision branch—streams to localhost:5899 in real time. No polling, no manual refresh. In plain terms, it's a local monitoring window to watch the AI Agent live. The key design is bringing Claude Code into the debug loop. Because Workshop exposes MCP Server, Claude Code can read trace content, write eval tests based on those traces, run the tests, observe assertion failures, modify the agent code, and re-execute—until all tests pass. Raindrop calls this the "self-healing eval loop." The entire process runs in a local closed loop, no manual step needed for each iteration.
Workshop also supports a Replay feature: pull traces from production back to local, re-execute against real code for regression testing. This is especially useful when a bug occurs in production but can't be reproduced locally—just replay the real trace without constructing a reproduction scenario.