Casebook MCP

A remote MCP server that turns AgentPostmortem, a public registry of documented AI-agent failures, into tools any agent can query. Deployable on Cloudflare Workers. Ships with a companion investigator agent built on the Claude Agent SDK that drafts postmortems grounded in real precedents.

Why: every team debugging an agent incident is rediscovering failure modes that are already documented. Casebook makes the registry agent-queryable, so Claude, Cursor, or any MCP client can ask "has anything like this happened before?" mid-investigation.

Tools

Architecture

MCP client (Claude Code, Cursor, agent SDK)
        |  streamable HTTP (JSON-RPC 2.0, stateless)
        v
Cloudflare Worker  src/index.ts (routing, per-IP rate limit)
        |          src/mcp.ts   (MCP protocol: initialize, tools/list, tools/call)
        |          src/search.ts (pure ranking and similarity logic, unit tested)
        v
src/data.ts: live agentpostmortem.com public API (/api/export, /api/search,
/api/tags) with a 5 minute in-memory cache, falling back to the bundled
dataset in data/ when offline.

The MCP transport is implemented directly against the 2025-03-26 streamable HTTP spec in stateless mode: a single POST /mcp endpoint handling initialize, tools/list, and tools/call. JSON-RPC batches are limited to 100 requests and rejected as a unit before dispatch when larger. No sessions, no Durable Objects, no auth (the data is public and read-only). A light in-memory rate limit (60 requests per minute per IP) keeps it polite.

tools/call accepts an optional JSON object for arguments; arrays, scalar values, and null return JSON-RPC -32602 Invalid params before tool dispatch.

Quickstart

npm install
npm test              # vitest: search ranking, tag filter, CSV parsing
npm run typecheck     # tsc --noEmit
npm run dev           # wrangler dev on http://localhost:8787

Smoke test the endpoint:

curl -s http://localhost:8787/mcp -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Add it to Claude Code:

claude mcp add --transport http casebook http://localhost:8787/mcp

Investigator agent

agent/investigate.ts uses the Claude Agent SDK query() API. Given an incident description, it connects to the MCP server, finds similar documented failures, pulls the top cases, and writes postmortem-draft.md. It runs on your local Claude Code subscription auth; no API key appears in the code.

# full investigation (uses the model)
npm run investigate -- "our support bot approved hundreds of fake refunds overnight"

# keyless mode for CI: stubs the model, still exercises the MCP server
npm run investigate -- --dry-run "runaway retry loop spammed customers"

Point it at a deployed server with CASEBOOK_MCP_URL=https://casebook-mcp.<account>.workers.dev/mcp.

Example transcript

$ npm run investigate -- --dry-run "our support bot was tricked by text in a ticket into approving hundreds of refunds"
Wrote postmortem-draft.md (1091 chars)

$ head postmortem-draft.md
# Postmortem Draft (dry run)

## Incident

our support bot was tricked by text in a ticket into approving hundreds of refunds

## Similar documented failures (source: live (cached))

### APM-0003: Cursor support AI hallucinates login policy, triggering mass subscription cancellations

In full mode the agent additionally calls get_case on the top precedents and produces a structured draft with suspected failure mode, contributing factors, and remediations grounded in the documented lessons.

Data

• Live: public read-only endpoints on agentpostmortem.com (/api/export for the corpus, /api/search for rich case detail, /api/tags).
• Bundled: data/cases.json holds 12 representative case files (prompt-injection refund exploit, runaway retry loop, stale-cache hallucination, tool-permission escalation, and more) used as an offline fallback and as the deterministic fixture for tests.

Deploy

npx wrangler deploy