Introducing Raindrop Triage
Today we're launching Raindrop Triage, an agent that investigates your AI agents.
Use it in Slack, on the web, or through MCP from your coding agent.
"The triage agent is great for the messy part of understanding agent behavior. You can ask quick ad hoc questions, spot broader patterns, and jump into the traces that show what's actually happening."
AI Engineer at Framer
"We have Triage running a Monday digest of our top issues and new patterns. It surfaces things we would've missed otherwise and points us to the relevant troublesome conversations."
AI Engineer at Speak
The problem with finding agent issues
A customer tweets that your agent is broken. Someone screenshots it into Slack. An engineer digs through traces for hours before posting: "yeah, the agent's making up order IDs when the lookup tool times out."
Before fixing it, they still need to find out how many people are affected and whether it's still happening. Triage can investigate those questions from Slack.
Triage in Slack
Tag @raindrop in a channel to ask a question or schedule a report:
@raindropevery Monday, summarize the biggest issues enterprise customers had@raindrophow's the rollout on flag checkout-v2?@raindropis this still happening? (with a screenshot of a customer complaint pasted in)
Triage replies in the thread and keeps context as you ask follow-ups. You can ask whether a fix shipped or an issue regressed, run experiments, and investigate customer complaints or Raindrop issues.
Triage through MCP
Your coding agent can call Triage through Raindrop MCP to investigate issues before working on a fix. For example, you can schedule Devin to ask Raindrop for the most important issue every night and try to fix it:
> Devin "check raindrop for the top enterprise issue and fix it"
→ Calling raindrop...
→ Found: list_projects tool fails when filtered by keyword
(43 conversations affected, started April 22)
→ Reading src/tools/projects.ts
→ Editing src/tools/projects.ts, removing keyword from tool schema
→ Running tests... ✓ all passing
→ Opening PR #847Claude Code, Cursor, Codex, and Devin can all use Triage through MCP. Your coding agent can read the failing production examples, make a fix, run tests, and open a PR.
Why we made our agent an MCP tool
We also expose Triage's underlying tools through MCP. Giving coding agents access to Triage itself lets them use what we've learned about investigating agent issues.
1. Task-specific harness
We've tuned Triage's harness to investigate production data, navigate conversations, and calculate how many users an issue affects. With the same tools, Triage generally investigates issues faster and more accurately than Claude Code.
Trying to tune for individual coding agents gave us little improvement, and we couldn't control their harnesses.
Making Triage an MCP tool lets us choose its model and harness while keeping it available to any coding agent that supports MCP.
2. Personalization
Raindrop learns which issues matter to you from the signals you refine, your feedback, the issues you fix, and your conversations with Triage. Your coding agent can use that context by calling Triage through MCP.
3. Context handoff
If you ask Triage to investigate a bug in Slack, you can then tag Devin to fix it. Devin can call Triage to pick up the investigation where you left off.