Blog
Long-form writing on AI coding agents, persistent memory, BYOK economics, and what we're learning building VibeKit.
-
AI-built apps by the numbers: speed, cost, and what people actually build
90 days of production data from an AI app builder: 2.7 minutes to a first AI-built version, 64 cents of model usage in a builder's first week, 78% built from an iPhone, and how much people iterate.
-
Can AI build an app for me? An honest map of what works in 2026
Yes, for a real class of apps. What AI app builders reliably deliver, the three places projects actually fail, and how to tell in one afternoon whether yours will work.
-
Can Claude build an app? Yes, and here is what that actually looks like
What Claude can and cannot build in 2026: the three ways to build an app with Claude, what it costs, whether it works on iPhone, and where projects actually fail.
-
Can Claude Code run 24/7? What it takes to keep a coding agent always on
Claude Code stops when your laptop sleeps. The four real ways to run a coding agent around the clock: local machine, VPS with tmux, cloud sessions, and managed hosting, with honest costs and failure modes.
-
How much does it cost to build an app with AI? Real numbers from 570 apps
Actual production data: the median AI-built app cost 17 cents of model usage over two months. Where the real costs are, what a message costs, and how AI compares to agencies and no-code tools.
-
Where to host AI apps: what an AI-built app actually needs to stay alive
Static host, serverless, VPS, or the builder's own hosting? What breaks when you export an AI-built app, why env vars are the usual casualty, and a checklist for choosing.
-
Can you run Claude Code on your phone? Yes — but where it runs matters
Use Claude Code from an iPhone or Android phone with Remote Control or a cloud session. Learn what runs on your machine, what needs to stay awake, and which setup fits.
-
OpenClaw hosting: five real ways to run it, from someone who runs hundreds
Where should you host OpenClaw? An honest comparison of the five real options — locally, Docker on a VPS, a Raspberry Pi, shared hosting, or a managed service — with the actual costs, failure modes, and maintenance each one carries. Written by the team that operates a multi-tenant OpenClaw fleet in production.
-
Loop engineering, explained by someone who runs the loop in production
Loop engineering is the shift from prompting an AI agent step by step to designing a system that prompts itself. Here's the ELI5, the five components, and the operational lessons from running thousands of these loops in production: the verifier matters more than the writer, the loop must close on production truth, and budget breakers are not optional.
-
Running a coding agent from your phone: three patterns and their tradeoffs
People are running Claude Code and Codex from their phones by three very different setups: on-device (Termux), a DIY remote box (VPS + Tailscale + tmux), or a hosted agent. Here's an honest breakdown of what each one actually costs you when the novelty wears off.
-
Does Claude Code support BYOK? (And Codex?) The honest answer
People search 'claude code byok' and 'does codex support byok' expecting a yes/no. The real answer is subtler: both CLIs always run on your own credentials — that's not a feature, it's the only mode they have. Here's what 'BYOK' actually means for Claude Code and Codex, the API-key vs subscription split, and the one thing the CLIs can't do that sends people looking.
-
How AI coding agents pick a model — and what BYOK changes
When you hand an AI coding agent a task, something decides which model runs it and whose account pays for the tokens. Here's the actual routing — bring-your-own-key direct vs platform-paid fallback — for Claude, GPT, and Codex, and how to pick the path that fits how much you build.
-
Can you BYOK on Replit, Cursor, Bolt, Lovable, or v0?
People search '[platform] byok' because they want to run the AI agent on their own Anthropic or OpenAI key instead of buying credits. Here's the honest, sourced answer for each major AI builder — including the trap where a platform says it supports 'BYOK' but only for AI inside the app you build, not the agent that builds it.
-
What an AI coding workspace actually costs: BYOK at 10, 100, and 1,000 calls a day
If you run a coding agent every day, the question isn't 'lifetime cost' — it's 'what's my daily burn?' Here's the real operational math for a BYOK AI workspace at three usage tiers, with measured per-call token costs, the markup the platforms don't show you, and the one factor that can zero the model bill entirely.
-
How to build an app from your phone in 2026
A practical, honest guide to building and shipping a real app from a phone — the four approaches that actually work, what each one is good and bad at, and how to pick. Covers AI app builders, Claude/ChatGPT, no-code, and persistent coding agents.
-
The AGENTS.md specification: every section, in priority order
There's no official AGENTS.md spec — this is the working one. Every section the file needs, the order that actually matters, the ~5KB budget that silently truncates it, and what to offload to TOOLS.md instead. Grounded in the template VibeKit regenerates for every hosted app.
-
Persistent agent memory — the three layers most tutorials skip
When you hear 'persistent agent memory' you probably think vector database. That's one layer of three, and it's the one most apps reach for first when they actually need a different one. Here's how we run persistence on VibeKit — workspace, session, and tool-tracked — and when each one is the right answer.
-
Running SEO for a real business with an AI agent — the playbook we use on vibekit.bot
We're a tiny team. Our SEO work — keyword research, baseline capture, query-intent audits, decision logs, blog drafts — is mostly driven by an AI agent reading a single instruction file. Here's the exact setup, what it does well, what we still do by hand, and the artifacts you can copy.
-
BYOK security: what an app can actually see when you paste your Anthropic key
Pasting sk-ant-api03-... into a 'bring your own key' app feels safe — your bill, your account, your control. The reality is more nuanced. Here's what a BYOK platform CAN see, what it SHOULD do, the attack scenarios that matter, and how to evaluate any vendor before handing over a key that bills against your card.
-
Why we picked AWS Fargate over WebContainers for an AI coding platform
WebContainers — the StackBlitz tech behind Bolt — boot in 2 seconds and feel magic. We picked the slower, heavier option (real Fargate containers) anyway. Here's why, and what it means for the kind of apps you can actually ship on the platform.
-
What persistent agents actually remember — and what they shouldn't
Every AI coding tool now claims 'persistent memory.' Most of them just have a longer chat window. Real agent persistence is a layered system: durable files, an indexed workspace, agent-side notes, and the AI provider's own context. Here's what we put where, and the few things we deliberately let the agent forget.
-
AGENTS.md: how it works, and why every AI coding agent needs one
AGENTS.md is the per-app contract every AI coding agent reads on every turn — not a system prompt. It scopes behavior, encodes architectural decisions, and survives context resets. Here's how it's built and why most agents are missing one.
-
BYOK vs platform-paid AI: the real cost of 1,000 features
Most 'all-in-one' AI coding platforms charge $20-40/mo and bake the AI cost into the subscription. Bring-your-own-key sounds nicer in theory. Here's the actual math on which model wins for which kind of user — with real per-task token measurements, not vendor-friendly approximations.
-
Why I'm starting this blog
Quick intro to what I'll write about: AI coding agents, persistent memory, BYOK economics, and the architecture decisions behind VibeKit. No 'top 10 tips' content.
-
Most 'AI agents' aren't agents — a definition that actually means something
The word 'agent' has been slapped on every LLM product launched in the past 18 months. Most of them aren't agents under any honest definition. Here's a tight three-property test, and how popular tools score against it — including VibeKit.