Raindrop Open-Sources Workshop, a Local MIT-Licensed Debugger That Lets Coding Agents Write and Run Their Own Agent Evals
Raindrop released Workshop, a free local debugger that streams an AI agent's tokens, tool calls, and spans to a browser and lets Claude Code write and fix evals against the trace.
Overview
Raindrop, an AI-agent monitoring startup, has released Workshop, which it describes as “an open-source, free, local debugger for AI agents,” according to the company’s Raindrop blog. The tool streams every token, tool call, and span an agent emits to a local browser dashboard and, distinctively, lets a coding agent such as Claude Code read those traces and write evaluation tests against them. The project’s GitHub repository is published under the MIT License and carries the tagline that it is “the local debugger your agent is missing.”
The first public release, v0.1.6, is tagged May 14, 2026 on the project’s GitHub releases page, matching the launch date given on the company’s blog.
What We Know
Workshop targets a specific pain point in agent development. “Debugging an AI agent is miserable,” the company writes on its blog, describing how failures hide inside nested spans and developers end up manually reading through thousands of spans to find what went wrong. Workshop’s answer is live local tracing: “Every span from your agent streams to your browser as it happens with 0 latency,” the blog states. The GitHub repository frames the same capability as “Live streamed traces. Every token, tool call, and span streams into Workshop as it happens.”
The tool runs as a local daemon with a streaming UI served by default on port 5899, and it persists traces to a local SQLite-style database file at ~/.raindrop/raindrop_workshop.db, according to the GitHub repository. Both the repository and the product page list curl -fsSL https://raindrop.sh/install | bash as the install command, per the Raindrop product page.
The feature Raindrop emphasizes most is what it calls a “self-healing eval loop,” in which a coding agent uses the captured trace to write and repair tests automatically. On its product page, the company describes the loop as: “Claude writes the eval, runs your agent, sees the failure, fixes the code, and re-runs — until every assertion passes.” The blog describes the same workflow as letting Claude “read the spans, write evals from the trace, and fix the code until the agent works,” exposing traces to the coding agent through a Model Context Protocol integration.
According to the GitHub repository, Workshop integrates with a broad set of agent frameworks and SDKs, including the Vercel AI SDK, the OpenAI Agents SDK, the Anthropic SDK, the Claude Agent SDK, LangChain, LangGraph, CrewAI, Mastra, Pydantic AI, DSPy, Google ADK, Strands, Agno, and Deep Agents. The same repository lists support for coding agents including Claude Code, Codex, Devin, Cursor, and OpenCode.
The project has iterated quickly since launch. The GitHub releases page shows ten tagged releases from v0.1.6 on May 14, 2026 through v0.1.15 on June 4, 2026, with the most recent adding the ability to import traces from the cloud.
Workshop builds on Raindrop’s existing commercial product. The company says the local tool uses “the same SDKs, the same schemas, and the same primitives” as its production monitoring system, according to the blog. Raindrop raised a $15 million seed round led by Lightspeed Venture Partners, announced December 1, 2025, according to the company’s press release. In that release, chief executive Zubin Koticha said, “Raindrop is the first monitoring platform to solve this problem” of detecting failures in long, autonomous agent trajectories, and Lightspeed partner Bucky Moore said of the team, “They defined monitoring for AI agents.”
What We Don’t Know
Raindrop has not published adoption metrics for Workshop, such as download or active-installation counts, in the materials reviewed here. Because the self-healing eval loop is demonstrated primarily with Claude Code, it is not clear from the company’s documentation how reliably the automated write-fix-rerun cycle performs across the other supported coding agents. The releases page shows rapid pre-1.0 versioning, and the project’s long-term stability and breaking-change cadence remain to be seen.
Analysis
Workshop slots into a fast-growing category of agent observability tooling, but it differs from cloud-hosted monitoring dashboards in two ways that matter to developers: it runs entirely locally, and it treats the coding agent itself as the consumer of the trace rather than a human reading a dashboard. By exposing traces over the Model Context Protocol, Raindrop positions the debugger less as a viewer and more as a feedback channel that lets an agent inspect and correct its own behavior. The MIT license and local-first design also let teams keep agent traces on their own machines, an option that contrasts with sending production telemetry to a vendor’s cloud — the same business in which Raindrop sells its paid product.