The Setup
We're a four-person AI studio in Seoul. Our production code at Hypemarc goes through Claude Code with a custom MCP server stack. After eight months of daily use, we have a set of patterns to share with fellow builders.
If you're considering Claude Code for your team, especially across a non-English working environment, this is what we learned.
The MCP Stack We Run
| MCP Server | Purpose | Where it pays off |
|---|---|---|
| Marblo (our own) | Task tickets, activity logs, status transitions | Every code change ties to a ticket automatically |
| GitHub (community) | PR creation, issue lookup, branch operations | Eliminates context switching to web UI |
| Linear (community) | Cross-team ticket sync | When clients use Linear, not Marblo |
| Custom domain MCP | Project-specific tools (deployments, log queries) | Per-client; we ship one for most engagements |
| Filesystem (built-in) | File ops with permission controls | Always on |
Five servers, configured per project. The Claude Code config lives in our repos so onboarding a new engineer is git clone + claude code ..
Pattern 1: Ticket-Driven Development
Every code change starts with a ticket in Marblo MCP. The Claude Code session opens, claims a ticket, works in scope, logs activity entries, and submits for review when done.
The main benefit is that the entire engineering history is queryable. Six months later we can ask:
"What was the reasoning behind the heterogeneous routing change in the marketing dashboard?"
…and pull the ticket's activity log, the related commits, and the design conversation. Without MCP-mediated ticketing this would require digging through Slack, GitHub PRs, and old Notion docs.
Pattern 2: Skill Files for Cross-Project Knowledge
Claude Code skills (folders under ~/.claude/skills/) hold our internal playbooks. We have:
seo-geo-full: install SEO + GEO foundation into any Next.js projectga4-full-tagging: drop full GA4 tagging into a projectmarblo-board-setup: scaffold a Marblo board for client engagementsinternal-deploy-flow: our Vercel + DNS + monitoring checklist
When we start a project, we invoke skills instead of repeating the same setup from memory. We open-sourced two of them (ga4-full-tagging-skill, seo-geo-skill). The rest are client-confidential.
Over time this adds up: a new engineer can be productive on day one because the team's accumulated patterns can be run as skills.
Pattern 3: Non-English Working Environment
This part is specific to teams working in a non-English language. We work in Korean. Claude Code handles this well, but there are two patterns that aren't obvious:
Pattern 3A: System prompts in English, conversations in Korean. In our experience, Claude follows instructions more precisely when system prompts are English. User messages and outputs in Korean are fine. We keep this split deliberately.
Pattern 3B: Code comments in the language of the team that maintains them.
Korean comments for internal code, English comments for code we open-source or share with global clients. The Claude Code skill comment-language-router enforces this automatically.
Pattern 4: Heterogeneous LLM Routing Inside the Editor
Claude Code defaults to Claude, but we route specific tasks to other models via MCP:
- Code refactoring: Claude Opus (in Claude Code natively)
- Boilerplate generation: Codex (GPT) via Marblo MCP (cheaper, equally good for known patterns)
- Image-grounded UI work: Gemini via custom MCP (best at multimodal)
- Long-context migration planning: Claude Opus 1M context (the new release)
The routing isn't manual. A Claude Code skill (route-by-task-type) decides which model handles which sub-task and reports cost back. In our use, engineering-side LLM costs went down without a noticeable drop in quality.
What We Stopped Doing
Lessons learned by getting it wrong first:
Stopped: keeping skills in our home directory only Skills only on a developer's local machine are tribal knowledge. We now version them in a private repo and sync via dotfiles.
Stopped: letting Claude Code commit without our pre-commit hook
Claude Code's commit defaults are fine for solo work. For team work, our pre-commit hook (linting, type-checking, secret scan) is non-negotiable. We removed --no-verify from any Claude Code config.
Stopped: running Claude Code without scoped permissions per project
The default permission mode is too permissive for client work. We use acceptEdits mode and require approval for destructive operations.
Stopped: assuming MCP servers stay up forever Community MCP servers can break with vendor changes. We pin versions and run health checks. A few times a community server broke, and the health check caught it quickly.
The Korea-Specific Edge
Working from Korea has practical advantages with Claude Code + MCP:
- Timezone arbitrage on incident response: when global services have late-night US issues, we're in our productive hours
- Korean-language Claude output is good: in our experience, Korean output quality is close to English
- Local MCP servers can integrate Korea-only tools (KakaoTalk, Naver, Toss APIs) that global community servers don't cover
We ship custom MCP servers for Korean integrations as part of most client engagements. Agencies outside Korea usually don't offer these.
What Changed for Us
Over eight months, our team noticed the following changes in everyday work rather than a controlled experiment:
- Time-to-first-PR for new engineers: noticeably shorter
- PR review cycles: fewer on average (Claude Code pre-flights catch more issues)
- Documentation freshness: docs stay current longer (because skills are runnable, they get updated when they break)
- LLM cost per engineering hour: lower after heterogeneous routing
What This Means for Your Team
If you're a small team (2-10 engineers) considering Claude Code + MCP, three things will determine whether it works for you:
- Are you willing to invest 2-3 weeks in setup? The first ROI shows up after the third client project, not the first.
- Do you have a place to put skills (private repo, dotfiles)? Without versioning, the patterns rot.
- Are you comfortable with heterogeneous LLM routing? Single-model setups don't capture the cost wins.
If yes to all three, the productivity ceiling lifts noticeably. If no, Claude Code is still useful, but the team effects don't materialize without the surrounding patterns.
Want to See It Working
We've onboarded several Korean teams onto a Claude Code + MCP workflow. The full setup including custom MCP servers takes about a week. Get in touch if you want a walkthrough on your stack.
Further Reading
- Building Your First Multi-Agent System with Marblo
- MCP Servers in Production
- Heterogeneous Agents in Production
Last updated: 2026-05-20. Our workflow keeps evolving as we learn.
Wrapping up
The Hypemarc Builder Community will keep sharing what we learn while building and operating AI agents, along with resources worth exploring. We'll continue gathering useful resources to share with fellow builders.
- Open chat: AI Builders
- Adoption advice and inquiries: Contact form
- Get the next post: Blog RSS · AI resource library