Community Notes

Claude Code vs Codex: Agentic Coding Gets Real

2025-12-05 • Homborsund AI

agents
claude
codex
convex
typescript

Agentic coding is no longer a demo—it's landing in real workflows. Tools like Anthropic's Claude Code and OpenAI's Codex CLI are racing to make “work with my codebase” a first-class capability: grounded conversations on repos, structured planning, and iterative execution with safety rails.

The shift is from autocomplete to agents: tools that plan, apply precise patches, and validate outcomes against your types and tests.

Øyvind on agentic CLI coding

“Claude Code was the first to make this feel real. There were earlier agent attempts, but Claude connected models, tools, and context engineering in a way that made CLI‑first development click for a lot of us.”

“You don’t really need a GUI. With text input + intent you get back working applications — fast. That unlocks a different way of building.”

Why this matters

The promise isn't just chat that can write functions; it's assistants that understand project intent, propose plans, change code surgically, and validate outcomes. Compared to earlier code completion tools, these agents can reason over broader context—tests, APIs, data models—and keep changes consistent with your architecture and type system.

How we build: type-safe velocity

In this repo, we lean on full TypeScript strictness and Convex for typed server functions. Pairing Convex's generated types with an agentic flow (via Codex) means we can move fast without losing safety: mutations and queries stay strongly typed end to end, and UI wiring catches mismatches early.

What we shipped that feels like a SaaS

  • Summit registrations: typed Convex mutations/queries with validation and storage.
  • Admin dashboards: list, edit, and delete registrations with server routes and SWR.
  • Tickets + verification: QR ticket pages and a verification flow for check-in.
  • Email notifications: transactional email hooks for confirmations/ops.
  • Auth gating: Google Workspace–protected admin routes.
  • Bot protection: Turnstile integration on public forms.
  • Multi-summit content: schedules, speakers, venue sections with reusable components.
  • Migrations + config: structured utilities and typed configs for repeatable changes.

In another world, this stack could be a handful of paid SaaS tools. Instead, we’re vibe-coding it: owning our data, iterating quickly, and keeping everything composable and type-safe.

Claude Code and Codex, head to head

The interesting bit isn’t who “wins” but how both product lines are converging on a similar idea: practical agents that plan, apply changes, and follow house style. Claude Code pushes hard on grounded reasoning and careful edits; Codex CLI leans into open, scriptable workflows you can tailor to your repo. Both are rapidly improving at handling multi-file refactors, incremental diffs, and safety checks.

“OpenAI’s Codex CLI jumped a level in the last month. In places it’s now on par — sometimes better. Codex Web still lags the CLI; I’m excited to see them meet the same high bar.” — Øyvind

Claude Code highlights

  • Repo-grounded reasoning and task-focused coding mode
  • Artifacts for inline previews and diffs
  • Strong natural language refactor guidance

Codex CLI highlights

  • Open, scriptable workflows with patch-based edits
  • Planning, validation hooks, and repo-specific context
  • Keeps changes minimal and reviewable

Try Codex locally

Install the terminal agent and point it at your repo:

npm i -g @openai/codex
codex

Also available via Homebrew: brew install codex.

Install Claude Code

Quick start

Install and launch from the terminal:

npm i -g @anthropic-ai/claude-code
cd your-awesome-project
claude

Editor/Desktop specifics vary — see the official guide: Claude Code setup.

Beyond SaaS: Build vs Buy in 2025

This isn’t just about our repo. Agentic tooling bends the SaaS calculus: many teams will spin up focused, owned platforms with today’s cloud primitives — whether that’s AWS, GCP, Azure, or (like us) Vercel + Convex — and ship features at a pace that once demanded multiple subscriptions. Authentication, email, admin panels, ticketing, analytics — all increasingly feel “in‑reach” for an in‑house stack powered by strong types and agents.

Cloud of your choice
Compose services on AWS/GCP/Azure or move fast on Vercel + Convex. Agents help wire the plumbing without sacrificing safety.
Security & ownership
Keep data residency and auth under your control. Adopt SaaS selectively where it truly compounds your velocity.

From user stories to intent stories

This may read like it’s only for developers, but the strength here is a crisp specification. “User stories” feel dated when agents can mine org and personal context. We’re moving toward “intent stories”: describe desired outcomes, constraints, and signals of success — then let agents propose plans, patch code, and validate.

  • Start with intent: a short brief with goals, constraints, and examples.
  • Plan: agents draft steps and surface dependencies and risks.
  • Patch: minimal, reviewable changes grounded in types and tests.
  • Iterate: keep a living spec; discard experiments freely and start fresh.

“The goal isn’t another backlog; it’s faster learning loops. Build, learn, throw away, repeat — as long as the spec lives on, we get better.” — Øyvind

What this unlocks for us

  • Faster feature loops: spec → plan → patch → validate in minutes.
  • Consistent quality: typed Convex endpoints and strict TS keep agents honest.
  • Operational clarity: agents propose changes; humans review, ship, and own outcomes.

Let’s discuss it live

This is an active area we’ll dig into face to face at our summits—tradeoffs, safety, and where these tools already pay off. Bring your questions and examples.

Sources