Home / Technology / See Your AI Agent's Every Move Locally
See Your AI Agent's Every Move Locally
15 May
Summary
- New tool offers local debugging for AI agents.
- It streams token, tool calls, and decisions in real-time.
- The tool features a self-healing eval loop for code correction.
Observability startup Raindrop AI has released "Workshop," a new open-source tool designed to debug AI agents locally. Launched today, this MIT-licensed tool functions as a local daemon and UI, streaming every token, tool call, and decision to a dashboard hosted on localhost. This allows developers to meticulously review their AI agent's actions, including any errors, and understand the root causes. The data is stored in a single, memory-efficient SQL database file.
A key feature is the 'self-healing eval loop,' which empowers coding agents to read traces, write evaluations, and autonomously fix broken code. For instance, if an agent fails to gather necessary information, Workshop captures the trajectory, enabling another agent like Claude Code to identify and correct the logic error, re-running until it meets defined assertions.
Workshop supports major programming languages like TypeScript, Python, Rust, and Go, integrating with popular SDKs such as Vercel AI SDK, OpenAI, Anthropic, and frameworks like LangChain and CrewAI. It is compatible with various coding agents, including Claude Code, Cursor, Devin, and OpenCode. Released under the permissive MIT license, Workshop aims to encourage community contributions and ensure enterprise data sovereignty, providing a more 'sane' local debugging approach for autonomous systems.
