Frontend Design Loop MCP

Coding agents can get a page functional. Frontend Design Loop makes it materially better with screenshot-grounded iteration and proof artifacts.

Use it when the base model got the page working but the result is still generic, flat, rough, or visibly under-designed. The main design workflow stays on one main provider and model lane by default, so multi-model routing is opt-in instead of the default story.

Quick Start

Install the current public build from PyPI:

pipx install frontend-design-loop-mcp

Set up every detected supported client:

frontend-design-loop-setup --install-all-detected-clients

Real MCP call example:

frontend_design_loop_design(
  repo_path="/absolute/path/to/site",
  goal="make the homepage look materially more premium without changing the information architecture",
  provider="gemini_cli",
  model="gemini-3.1-pro-preview",
  preview_command="python3 -m http.server {port}",
  preview_url="http://127.0.0.1:{port}/index.html"
)

Related MCP server: vibe-annotations

Agent-Owned Toolkit Variant

This repo now also ships an additive MCP entrypoint for agent-owned frontend loops:

• console script: frontend-design-toolkit-mcp

• module fallback: PYTHONPATH=src .venv/bin/python -m design_toolkit.server

That variant is intentionally narrow:

• get_playbook

• build_context

• run_gates

• preview_start

• capture_screenshots

• preview_stop

It does not hide vision scoring, creativity judging, or patch generation behind MCP. The host agent owns planning, edits, screenshot review, scoring, iteration, and winner selection directly.

See test-prompt-codex.md for the end-to-end Codex exercise prompt.

What The MCP Does

frontend_design_loop_design is the main workflow:

• the host agent points the MCP at a real repo plus a concrete design goal

• the MCP boots a local preview, captures screenshots, and iterates against the rendered result

• the same main provider and model lane is used by default across planning, generation, and vision unless you explicitly override it

• the MCP returns the winning patch plus screenshots and run artifacts

frontend_design_loop_eval is the proof workflow:

• use it when the host agent already has the patch

• the MCP applies the patch in an isolated worktree, runs deterministic checks, captures screenshots, and returns proof artifacts

This is the wedge:

• coding agents can already get pages working

• this MCP helps them make pages materially better

• screenshot-grounded iteration plus proof artifacts is the differentiator

Official MCP Registry metadata is tracked in server.json.

Proof Gallery

The public proof set uses owned/generated GA SMB previews plus the ACA full-page before/after.

Selected Hero / Top Crops

ACA Full-Page Before / After

Before: early ACA full homepage.

After: rebuilt ACA homepage with a stronger hero, cleaner sequencing, and a materially better full-page result.

See the proof notes in the case studies index.

How It Works In Practice

1. Point the MCP at a real repo and give it a concrete design goal.

2. It creates an isolated worktree, boots a preview, and captures rendered screenshots.

3. It iterates against the actual rendered page instead of only raw code.

4. It returns the winning patch, screenshot proof, and run artifacts so the host agent can judge the result.

Workflow Summary

frontend_design_loop_design

Use it when:

• the page is functional but weak

• the section structure is there but the design is not

• you want the MCP to improve the page instead of only judging it

Key defaults:

• one main provider + model lane by default

• planning_mode="single"

• vision_mode="on"

• section_creativity_mode="on"

• split planner or vision lanes only happen when explicitly requested

frontend_design_loop_eval

Use it when:

• the host agent already has the patch

• you want deterministic checks, screenshots, and artifact capture

• you want the host agent to judge the result from returned screenshots

Returned proof fields include:

• deterministic_passed

• vision_pending

• vision_scored

• final_pass

• run_dir

• candidate_dir

• screenshot_files

• patch

frontend_design_loop_solve

frontend_design_loop_solve still exists for advanced unattended workflows, but it is not the main public story.

Install And Setup

Public install now

pipx install frontend-design-loop-mcp
frontend-design-loop-setup --install-all-detected-clients

GitHub install remains the fallback:

pipx install git+https://github.com/alexalexalex222/frontend-design-loop-mcp.git
frontend-design-loop-setup --install-all-detected-clients

Local clone path

git clone https://github.com/alexalexalex222/frontend-design-loop-mcp.git
cd frontend-design-loop-mcp
./scripts/setup.sh

The local setup path:

• creates .venv

• installs the package

• installs Playwright Chromium

• installs detected client entries when supported clients are present

• runs the built-in doctor

• runs the stdio smoke test

If you want the repo-local environment without auto-installing client entries:

FDL_SKIP_CLIENT_INSTALL=1 ./scripts/setup.sh

Setup helpers

Bulk installer:

frontend-design-loop-setup --install-all-detected-clients

Targeted installers:

frontend-design-loop-setup --install-claude --scope user
frontend-design-loop-setup --install-codex
frontend-design-loop-setup --install-gemini
frontend-design-loop-setup --install-droid
frontend-design-loop-setup --install-opencode

Config printers:

frontend-design-loop-setup --print-claude-config
frontend-design-loop-setup --print-codex-config
frontend-design-loop-setup --print-gemini-config
frontend-design-loop-setup --print-droid-config
frontend-design-loop-setup --print-opencode-config

Safety Defaults

• custom commands are parsed as shell-free argv by default

• shell syntax, substitutions, and inline interpreter execution like bash -c, python -c, and node -e require unsafe_shell_commands=true

• preview_url must match the launched local preview origin and port by default

• external preview fetches require unsafe_external_preview=true

• preview readiness checks reject cross-origin redirects, and browser screenshots block cross-origin subresources by default

• auto-context skips common secret-bearing paths such as .env*, .git/, .aws/, .ssh/, .config/gcloud/, .docker/, .kube/, token-named files, and service-account-style JSON

• native CLI providers inherit a minimal allowlisted environment instead of the full host shell environment

• shared worktree reuse directories are off by default

Client-side vision is the default proof path for frontend_design_loop_eval, so the host agent can judge the screenshots without provider credentials.

Proxy-only MiniMax vision lanes are explicitly treated as structural-only review:

• vision_review_mode="proxy_structural"

• they do not count as full automated visual scoring

Verification

Offline preflight:

PYTHONPATH=src .venv/bin/python scripts/preflight_check.py

stdio smoke:

PYTHONPATH=src .venv/bin/python scripts/smoke_mcp_stdio.py

Built-in doctor:

frontend-design-loop-setup --doctor
frontend-design-loop-setup --doctor --smoke

Docs

• Workflow reference

• Launch checklist

• Directory submission copy

• Case studies

Distribution State

Current public install path:

pipx install frontend-design-loop-mcp