[nemolize]

What happened

The Durable Objects free tier (100,000 rows_written per day) was exceeded on 2026-11-05 UTC, and writes began failing with Exceeded allowed rows written in Durable Objects free tier. Traffic was a steady 500-600 requests/min of GET/POST /sse from Claude.ai's connector.

Cause

Every GET /sse opens a new MyMCP Durable Object, and the agents library creates its bookkeeping tables on each new object. Measured in the Workers test pool, one new GET /sse costs 54 SQL statements writing 50 rows, plus 6 storage.put calls.

Fix

• /mcp and POST /sse now run without a Durable Object: a fresh MCP server per request, no Mcp-Session-Id. No tool sends server-initiated messages, so nothing depends on a session.
• A streamable client's follow-up GET /sse (it carries mcp-protocol-version) and DELETE get 405 instead of reaching a Durable Object. GET /mcp also gets 405.
• The legacy SSE handshake (GET /sse without that header, then POST /sse/message) is unchanged and still uses a Durable Object.

Result

• POST /mcp and POST /sse (initialize, tools/list, tool calls) construct 0 Durable Objects; a new E2E test asserts this, with a legacy GET /sse as a positive control that constructs exactly 1. The full suite passes.
• Not deployed, so production rows_written is not measured yet.
• Trade-offs: a client that only speaks legacy SSE still costs one Durable Object per connection, and its reconnect after a dropped stream now gets 405 (its headers match the streamable follow-up GET). Both are stated in the README.

Test plan

• After deploy, compare rows_written per request via the GraphQL durableObjectsPeriodicGroups sum.rowsWritten.
• After deploy, check cpuTime for initialize, tools/list and tools/call in Workers Observability: each request now registers every tool, and a local Node benchmark put tools/list near 8-14 ms against the Free plan's 10 ms limit.
• Confirm the Claude.ai connector still connects and lists tools.