Claude Code agency — production software, senior review, 40-60% off traditional rates.

A Claude Code agency is a development studio that delivers production software primarily through Anthropic's Claude Code CLI, paired with senior engineering review on every commit. The team writes a spec, lets the model execute against it under a strict skills + hooks + memory harness, reviews every diff before it ships. The output rate is 3-5x a traditional shop on well-defined scope; the review rigour matches a senior pair. Wrong fit for ambiguous discovery — those projects need consulting first, then Claude Code execution.

Three differences. Speed — well-defined features ship in days not weeks because the model handles the boilerplate, the engineer handles judgment. Cost — 40-60% under traditional agency rates for equivalent scope because billable hours drop. Discipline — every project ships with a skills library, hooks, and a memory system that future sessions inherit, so the work compounds instead of resetting each engagement.

Yes if the harness is right. Three guarantees on every project. (1) Hooks block direct production writes — settings.json gates require human confirmation. (2) Tests written before features, run on real CI. (3) Senior engineer reviews every commit before merge. None of those are optional. Without the harness, vibe-coded output fails in production; with it, the failure rate is comparable to traditional engineering and faster to remediate when it does happen.

Four bands. Feature build (1-2 weeks): 5,000-15,000 USD. Product surface (4-8 weeks): 15,000-45,000 USD. Platform build (12-20 weeks): 45,000-150,000 USD. Programmatic SEO site (variable scale): 25,000-90,000 USD depending on page count and data complexity. Roughly 40-60% under traditional agency rates because the model carries the boilerplate; the human-review premium is real but bounded.

Both, with split responsibility. The model writes the boilerplate, the repetitive transformations, the test scaffolds, the typed shapes. The engineer writes the spec, picks the architecture, reviews every commit, makes judgment calls on edge cases, ships the final review. Anyone selling 100% autonomous output is shipping unsafe code; anyone selling 100% human output is overcharging. Both numbers are real and split roughly 60/40 model/human in line counts, 30/70 model/human in decision weight.

Yes — feature builds are the most common engagement shape. The minimum useful scope is roughly one week of work (5,000-15,000 USD). Below that, the discovery overhead consumes the saving. Above it, the agency model compounds quickly because skills and memory accrue across the engagement.

Three patterns. (1) Project-specific Claude Code skills lock down conventions — testing framework, deploy command, file structure, naming. (2) Memory files in .claude/ capture quirks that persist across sessions ("never edit X without Y", "Y deploys via Z"). (3) Hooks in settings.json refuse to let the model run destructive commands or push without preview deploy. Result: the agent acts the same on session 30 as session 1.

Primary stacks. Next.js + React 19 + Server Components. Astro with content collections or headless CMS. Supabase Postgres + Auth + Storage. Vercel and Netlify hosting. Tailwind + shadcn/ui for UI. Stripe for billing. Anthropic Claude API for AI features. Adjacent: Payload CMS, Sanity, WPGraphQL, Cloudflare Workers. Out of scope: Java, .NET, native iOS / Android, embedded, anything we cannot defend in code review.

Yes — handoff includes both the code and the Claude Code setup that produced it. Your team inherits: project-specific skills, settings.json hooks, memory files in .claude/, documentation generated during the build, plus any CI / deploy / SEO linter scripts written for the project. The next person to open Claude Code on the repo starts with the same context we did. The work compounds; it does not lock you in.

Three reasons. Twelve years of senior engineering practice — Claude Code is a tool layered on real production experience, not a substitute for it. Twelve thousand WordPress sites shipped at Seahawk Media before the AI tooling era — same engineering judgment applied to a faster delivery shape. And this site is the proof — Astro + Supabase + Vercel, built with Claude Code, with a working LLM citation tracker, programmatic SEO, schema correctness, and an admin layer that survives an Ahrefs audit. If you want a Claude Code agency that ships the work it talks about, this is it.