Category
Practical guides to building AI-powered software: MCP servers, agent tooling, LLM integrations and making apps readable by AI assistants.
AI Engineering
Build a ComfyUI custom node with its own JavaScript UI: INPUT_TYPES and the V3 schema, WEB_DIRECTORY, passing UI picks into node outputs, routes, saving state and publishing to the Comfy Registry. Tested on ComfyUI 0.42.1.
AI Engineering
Tested setup for adding a remote MCP server to Cursor, Claude Code, VS Code and Claude Desktop, using a real public endpoint. Copy-paste configs, curl and Inspector checks, mcp-remote, and what tools/list costs in latency and tokens.
AI Engineering
How to move LLM calls out of request handlers and into a BullMQ queue: idempotent jobs, Retry-After aware backoff with jitter, worker rate limits, schema-validated model output and a dead-letter queue. Tested TypeScript code.
AI Engineering
How Benchline runs 61 browser tools where each one is a pure function shared by a web page, a stateless Streamable HTTP MCP server and a JSON API, using real code from the repo.
AI Engineering
I upgraded a real AI SDK 6 app to AI SDK 7. The codemod fixed 15 lines in 20 seconds but only 1 of 9 broken tests. Here's every step, the behaviour changes TypeScript misses, and the OpenTelemetry trap.
AI Engineering
Both official MCP SDKs had a high-severity OAuth flaw (CVE-2026-104850). I reproduced it locally, found that upgrading alone doesn't fix bundled providers, and show how a production MCP server handles OAuth, per-tool scopes and audit logging.
AI Engineering
I built the same service as a REST API and as a stateless MCP server, then measured latency, throughput, bytes and the tokens a model reads. Here is when you need MCP, when a plain API wins, and how to wrap one with the other.
AI Engineering
I loaded 100,000 OpenAI embeddings into pgvector 0.8.7 on Postgres 18 and Qdrant 1.19 and measured recall, latency, index build and filtered search, then compared Pinecone's documented limits and pricing. When you need a vector database, and when Postgres is enough.
AI Engineering
MCP is stateless since the 2026-07-28 spec: no initialize, no Mcp-Session-Id. What changed on the wire, a tested TypeScript SDK v2 server on two instances, handles, requestState, MRTR elicitation and a migration checklist.
