Agentic Development Workflows
Claude + MCP
Claude as a teammate with hands: websites, repos, databases, docs.
01
Problem
Chat-only AI has a copy-paste problem: it can reason about your code, your data, and your docs, but it can't touch them. Every insight dies in the clipboard, and context evaporates between sessions.
MCP changes the contract — Claude connects to servers that expose real tools with real permissions. The question worth answering: what does a marketing-engineering workflow look like when the AI can actually do the work?
02
Solution
I run Claude with a set of MCP servers as a standard working environment: filesystem and GitHub for code, Postgres for data, browser automation for testing, and docs tooling for knowledge management.
This portfolio is itself an example — built end-to-end by Claude Code: scaffolding, design system, copy, components, and deployment checks. The workflows below are the ones I use weekly.
Capabilities
What it does
Building websites
Scaffold, design, and ship complete sites — Claude writes the code, runs the build, opens the browser, and fixes what it sees.
Editing repositories
Multi-file refactors and features with the filesystem MCP: Claude reads the codebase, makes coordinated edits, and runs the tests.
Working with GitHub
Issues triaged, PRs opened and reviewed, CI failures diagnosed — through the GitHub MCP instead of tab-switching.
Reading databases
Plain-English questions against Postgres: Claude writes the SQL, runs it read-only, and explains the result with the schema in context.
Managing documentation
Docs that drift from reality get regenerated from the source of truth — Claude diffs behavior against docs and proposes updates.
Rapid prototyping
Idea to clickable prototype in an afternoon: Claude scaffolds, iterates on feedback, and screenshots its own work for review.
Architecture
How it's put together
The shape of every workflow: Claude in the middle, MCP servers as typed tool surfaces, real systems behind them. Permissions live at the server layer, so the AI gets capability without getting the keys.
Interface
Claude Code
CLI + editor
Claude Desktop
Protocol
Model Context Protocol
Typed tools · resources · prompts
MCP Servers
Filesystem
GitHub
Postgres
Read-only
Browser
Preview + test
Docs
Systems
Repos & Sites
Databases
CI / Deploys
Knowledge Bases
Perspective
Why this matters for marketing teams
Marketing runs on systems marketers can't usually touch: the CMS codebase, the analytics warehouse, the ad platform APIs. Every change waits in an engineering queue. Agentic AI with MCP collapses that queue — the person with the marketing judgment can now safely operate the systems.
That's the thesis of my whole toolkit: the highest-leverage marketer on a modern team is the one who can go from 'we should test this' to a shipped test without a handoff. Claude with MCP is how I work that way daily.
Challenges
What was hard, and what I did about it
Permissions worth trusting
Giving an AI write access is an engineering decision, not a vibe. Database MCPs run read-only, deploys require approval, and anything destructive stays human-confirmed — capability scoped at the server, not the prompt.
Context discipline
Agents drown in their own context. Effective sessions front-load a tight brief — goal, constraints, definition of done — and use project memory files so the next session starts warm.
Verification as a habit
The workflow only compounds if the agent proves its work: run the build, execute the tests, screenshot the page. 'It should work' is banned; 'here's it working' is the standard.
Tech stack
Built with
Outcomes
What changed
- This sitewas built end-to-end with Claude Code + MCP — the proof renders in your browser
- Hours → minutesfor prototypes, refactors, and data questions that used to wait on handoffs
- 0 clipboardsinsights become actions in the same session they're generated
Next project
AI View Sync