I work with four tools (actually more, but let’s not overload ourselves). They look like they overlap. They don't, much. Each one owns a different layer of the stack. The rest of this post is what each one is.
Softr is a full-stack no-code platform. You bind blocks (List, Details, Form, Chart, Kanban) to a data source. Airtable, Google Sheets, Supabase, Postgres, REST, or Softr's own database. You get a working web app with auth, roles, record-level permissions, and gated content. The 2025/2026 generation added native Workflows (visual triggers, conditions, branches, AI steps), an AI Co-Builder that turns a prompt into a working app, and Database AI Agents that enrich records per-field on their own. The ceiling is layout. Blocks stack, period. No pixel-perfect, no free-form canvas. SSO is paywalled at Enterprise. Performance degrades past about 10,000 visible records. I use Softr where security matters more than novelty, and where I want non-engineers to extend without breaking it.
Lovable is a prompt-to-React generator with two-way GitHub sync. You describe a site or an app. Lovable writes React + TypeScript, styled with Tailwind + shadcn/ui + Radix, and commits to a real Git repo you own. Newer projects can ship as TanStack Start with SSR. kenove.no is one of them. Open the same repo in Cursor or Claude Code CLI, edit, commit, and Lovable picks up your changes on the next session. Supabase is wired by default. Agent Mode explores the codebase, reads logs, runs web searches, and auto-fixes its own loops. The ceiling: bug-fix loops burn credit, async timing in Supabase flows is finicky, and code quality degrades as the file count grows. The bilingual rebuild you're reading was Lovable and Claude Code in handoff.
Claude desktop is the thinking room. macOS and Windows. Drop in a PDF, an arXiv paper, a folder of notes. Cross-session memory carries context forward, rolled out to the free tier in March. Projects hold persistent local folders and custom instructions. The 2026 update added Cowork for autonomous local tasks, Computer Use to drive mouse, keyboard, and screen on Pro and Max, Live Artifacts that refresh on open, and 20MB of persistent storage per artifact. MCP servers are configured via claude_desktop_config.json or one-click .mcpb bundles. I use desktop for long-form reading, drafting, and anything that benefits from a GUI agent over a terminal one.
Claude Code CLI is the terminal-native agentic loop. Read, edit, write, run commands, search, follow a plan across many steps, with file checkpointing so any change can be reversed. Plan Mode (Shift+Tab twice) locks the agent into read-only analysis before execution. Hooks fire at lifecycle events to enforce policy and run validation. Skills add domain capabilities as small file-based markdown bundles. Subagents dispatch isolated workers in parallel. Explore for search, Plan for analysis, general-purpose for research, plus custom agents. MCP servers extend the tool surface. In this project, that means Sanity, Gmail, Drive, Calendar, Vercel, OmniSocials, and a handful of others. I use Claude Code for the codebase: building, refactoring, migrating, automating. The post you're reading was edited from a terminal via Claude Code.
The lanes blur, on purpose. Lovable can write code Claude Code could write. Claude Code can draft copy Claude desktop could draft. Softr can host content Lovable could ship. The question I ask isn't which tool can do the job. It's where the work belongs. A signed-in client portal belongs in Softr. A teaching publication belongs in Lovable plus Claude Code. A research session belongs in Claude desktop. A migration touching forty files belongs in Claude Code CLI.
None of this has to be paid-max. The free tiers do real work. The lift comes from knowing which tool owns which problem.
More on each, with the actual prompts and files, in the walkthroughs.
More notes like this at kenove.no. Subscribe at the bottom if you want them when they land.