Casebook Chat

A streaming AI chat UI that investigates AI-agent failures. Ask about an incident and the assistant searches the live AgentPostmortem case registry, pulls full case files, and answers with cited case IDs. Tool calls render inline as collapsible chips while the response streams.

Part of a three-project portfolio, alongside RelayG and hire.agentpostmortem.com.

Architecture

Browser (Vite + React, @ai-sdk/react useChat)
   |  POST /api/chat (UI message stream)
   v
Cloudflare Worker (Hono + Vercel AI SDK streamText)
   |                         |
   v                         v
Groq (llama-3.3-70b)     MCP registry (JSON-RPC tools/call)
                         https://mcp.agentpostmortem.com/mcp

One Cloudflare Worker serves both the static UI (Workers assets) and the API. POST /api/chat runs streamText against Groq with three tools defined via the AI SDK tool() helper and zod schemas:

• search_cases(query) - full-text search over failure case files
• get_case(id) - full detail of one case (facts, unknowns, lessons)
• similar_failures(description) - match a described incident against the casebook

Each tool is a fetch to the MCP server using JSON-RPC tools/call. Multi-step tool chains are allowed (up to 5 steps), and the result streams back with toUIMessageStreamResponse(), so the UI sees tool inputs and outputs as typed message parts.

Quickstart

npm install
cp .dev.vars.example .dev.vars   # add your GROQ_API_KEY (free at console.groq.com)
npm run build                    # build the UI into dist/client
npm run dev                      # wrangler dev on http://localhost:8787

Without a key, /api/chat returns a friendly JSON error that the UI displays; the shell still loads.

Other commands:

npm run check    # tsc --noEmit for web and worker
npm test         # vitest (JSON-RPC client + UI helpers)
npm run deploy   # build + wrangler deploy (set GROQ_API_KEY as a Worker secret)

Honest notes

• Inference runs on the Groq free tier, so rate limits and occasional slowdowns apply.
• The registry data is community-documented and comes straight from the live MCP server; the assistant is instructed to cite case IDs and to say so when the casebook has nothing relevant.

Links

• Registry MCP server: https://mcp.agentpostmortem.com
• Hiring page: https://hire.agentpostmortem.com
• RelayG: https://github.com/AgentPostmortem/relayg