Dora Noda
Software Engineer
2313 posts · View all authors
Your Deploy Agent's Sandbox Will Die Mid-Rollout — Here's How Its MCP Server Should Survive That
MCP's July 2026 spec update made the protocol stateless — which means the burden of surviving a crashed sandbox now falls entirely on the MCP server. Here's the Postgres checkpoint design, borrowed from LangGraph, that a deploy-from-chat PaaS needs to make a multi-step rollout outlive the agent that started it.
Designing upgrade_cluster: An MCP Tool Contract for Fleet-Wide Kubernetes Node Upgrades
A deploy/rollback MCP tool that goes wrong costs a bad release. An upgrade_cluster tool that goes wrong can leave a fleet straddling a Kubernetes version-skew boundary mid-drain. Here's a concrete tool contract — plan mode, per-node health gates, and a hard stop — for handing an agent that authority safely.
The Rollback Floor: Designing a Deploy-From-Chat MCP Server That Can't Roll Back Too Far
A dashboard's rollback button only ever offers 'go back one.' An MCP tool wrapping the same API can't assume that restraint — here's a concrete tool contract, server-side floor, and audit schema that stop an agent from rolling back past the last known-good release.
MCP's '10,000 Connections at Sub-50ms' Claim Is True — and Doesn't Mean What You Think
The 2026 benchmark headline about MCP servers handling 10,000 connections at sub-50ms is technically true and mostly beside the point. Here's what a production deploy-from-chat MCP server actually needs when a fleet of agents, not one chat session, is calling it.
MCP's Tasks Extension Is Built for Work That Outlives a Single HTTP Request — Exactly the Shape of a Deploy
MCP's July 2026 spec ships a Tasks extension so a deploy tool call can outlive a single HTTP request — here's how tasks/get polling beats the hand-rolled deploy_status workaround every deploy-from-chat integration had to invent before it.
What Self-Hosting OpenMeter's Kafka+ClickHouse Stack Actually Costs a Git-Push PaaS
OpenMeter's self-hosted stack adds two new stateful services — Kafka and ClickHouse — to bill per millisecond. Here's what that actually costs in Hetzner hardware and ops burden against billing tenants on signals a PaaS already tracks for free.
Railway's Three Years Without a Price Increase vs Render's April 2026 Repricing: What Sticker Stability Actually Signals
Railway hasn't touched its plan fee or rate card since 2023; Render restructured its entire workspace pricing in April 2026. Here's what the two timelines actually reveal about cost structure, growth strategy, and why a flat sticker price isn't the same claim as a flat bill.
Render, Railway, and Fly.io's Hiring Pages: Reading Job Counts Before You Migrate
Render lists 40 open roles, Railway 6, Fly.io zero. A framework for reading a hosted PaaS vendor's own careers page and product changelog as a leading indicator before you commit to it — or migrate off it.
Wassette Runs MCP Tools as WebAssembly, Not Containers: What That Actually Buys a Self-Hosted Fleet
Microsoft's Wassette runs MCP tools as WebAssembly Components with deny-by-default WASI permissions instead of a container or a Firecracker microVM. Here's what the microsecond instantiation and policy.yaml grants actually buy a self-hosted fleet — and the GPU access, fork(), and resource-quota gaps that still mean a heavier sandbox for everything else.