ok this blew up more than expected. had fun building it tbh. the animation is just for the demo. strip it out and it's a self-hosted email client on @vercel ai gateway + jev + gmail api. repo: github.com/fazlerocks/jevmai…

jev is INSANE for email. sorted 1,000 gmail messages in 76 seconds. cateogies: needs reply, updates, promos, sales, spam. parsed every single one. total cost $0.03. might just make this my daily mail client now. open source, link below @typesafeai

repo: github.com/fazlerocks/jevmai…

jev is jevons paradox in real time small decisions (yes/no… pick one… score this) just went from cents and secs to nothing and ms every not-worth-an-llm-call idea is worth it now we should see inference spend go up drastically, not down. in the next few weeks

New in DearAgent: ADDRESS_MATCH_PATTERN. Set a regex like ^agent- and only those addresses become agent inboxes. Everything else forwards to your normal mailbox. thanks for the PR @xvks42! 🫡

open-source AgentMail alternative on Cloudflare Workers. dearagent.sh

One of my favorite features of Darkshift is the ability to preview the agent's work on localhost without setting up local dev. After a task finishes, the factory emits a one-liner: npx darkshift preview <token> --port <port> The CLI spins up a short-lived dedicated sandbox with the exact PR state (dev server already running), reverse-proxies it to localhost with full WS/HMR passthrough, opens the browser, and tears the sandbox down cleanly once you're done. You don't need to change your workflow. If you've been verifying your work on localhost, you can still test the agent's work locally without ever pulling the code to your local machine.

adoption for v0, replit, and lovable has only gone up in the last few months. but as products built with these tools scale… you can't go prompt to production without verifying what actually works. so we're taking the ADLC further. a software factory builds the next feature and leaves a PR ready to ship with a verifiable QA video attached. if the QA fails, it loops and keeps working until it passes. Spotify, Uber and Ramp already spend quarters building factories like this internally. high-growth startups can't afford that. we built darkshift for ourselves and have shipped a lot of features with it at Hashnode and Bug0 over the last 8 weeks. now it has a name and a website. it's invite-only for now. i'll waive the platform fee for 12 months in exchange for feedback.

Introducing Darkshift - a software factory we’ve been using to ship Hashnode + Bug0. > multi harness: claude code, codex, grok build or kimi > BYOK > web, slack or linear > 100% cloud dev > QA-first with evidence-backed PRs > Multiplayer darkshift.dev

my context window today

Everyone is suddenly building an AI software factory: a system where coding agents take a ticket to a verified pull request with no human in the middle. The term is from Hitachi in 1969, and this is its third revival. What strikes me about the category is how converged it already is. Devin, Factory, 8090, and FactoryKit all run the same five stages: intake, isolation, implementation, verification, PR. When four vendors independently arrive at one architecture, that usually means it's durable. What's missing is measurement. There is no public benchmark for ticket-to-merged-PR reliability across any of these products. The naming is outrunning the measuring, and buying decisions are running on demos. Full disclosure, we build one of the four, so I wrote this with the gaps in as much detail as the pitch: hashnode.com/blog/ai-softwar…

wrote a blog comparing the best Postman alternatives in 2026 -- what they do well, where they fall short, and which teams they're actually built for. 👇

Hashnode now has an official skill for AI agents. npx skills add Hashnode/gql-skill That's it. Claude Code, Cursor, or Copilot can publish to your blog, edit old posts, manage drafts, upload cover images. Why a skill instead of just pointing agents at our GraphQL docs? Because agents kept getting the details wrong. They'd paste access tokens into chat, retry calls that were never going to succeed, ask a feed for 500 posts when the API caps a page at 100. The skill bakes those rules in, so the first request works. Write your draft wherever you write. Then: "publish this to my Hashnode blog." github.com/Hashnode/gql-skil…

QA has reached new heights.