Claude Code with preset AI

Use Claude Code with preset AI as a governed design-system runtime.

Setup

One command:

claude mcp add --transport http preset https://mcp.presetai.dev/mcp

OAuth runs on first request. Verify the connection:

claude mcp list

You should see preset with status connected. The full per-tool install matrix lives in Connect an AI coding tool.

Recommended operating loop

The four-step loop that keeps Claude Code aligned with the system:

Read context first: inspect relevant presets, patterns, and tokens; fetch active constraints for the target file or intent
Generate with constraints: request code by preset name and semantic intent; avoid hardcoded values and unapproved primitives
Validate before and after: run pre-generation checks when available; run generated-code validation and drift checks
Propose before apply (for system-level changes): for new or changed presets or rules, use the proposal lifecycle; apply only after validation and required approvals

Prompting pattern

Structure your prompts like this:

Intent: what you are building
Context: target route or component, and constraints
Required artifacts: preset names, token categories, patterns
Validation ask: request drift and compliance checks before finalizing

Example:

Build a destructive confirmation flow for account deletion. Use approved destructive presets, token-based colors, and the interaction rules from the design system. Validate generated code and report drift and compliance findings before final output.

Read, generate, apply

Read

at the start of a task
when uncertain about naming or policy
when working in unfamiliar domains or components

Generate

only after retrieving required constraints or artifacts
for implementation and refactor suggestions

Apply

direct apply for low-risk local code changes with passing validation
proposal + approval path for canonical system artifacts

Common failure patterns

generating from generic UI patterns without querying preset AI first
hardcoding style values where tokens exist
skipping post-generation validation
applying system-level changes without the proposal workflow

Team policy

Adopt a simple policy:

No merge without passing drift and compliance checks
No canonical model writes from AI without a proposal trail
Require explicit approval in stricter compliance modes