Frontend Design Loop MCP README
The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Frontend Design Loop MCP listing page.
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:
Set up every detected supported client:
Real MCP call example:
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_playbookbuild_contextrun_gatespreview_startcapture_screenshotspreview_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
 11 Budget Movers Augusta |  13 Peachtree Flooring Atlanta |  19 TNT Cabinets Columbus |
 21 Henry Plumbing Savannah |  22 Silverback Electric Savannah |  25 Robins Body & Paint Warner Robins |
 34 Proof Roofing Services Gainesville |  45 Metro Storage Columbus |  47 Miller Light Construction Commerce |
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
- Point the MCP at a real repo and give it a concrete design goal.
- It creates an isolated worktree, boots a preview, and captures rendered screenshots.
- It iterates against the actual rendered page instead of only raw code.
- 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+modellane 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_passedvision_pendingvision_scoredfinal_passrun_dircandidate_dirscreenshot_filespatch
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
GitHub install remains the fallback:
Local clone path
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:
Setup helpers
Bulk installer:
Targeted installers:
Config printers:
Safety Defaults
- custom commands are parsed as shell-free argv by default
- shell syntax, substitutions, and inline interpreter execution like
bash -c,python -c, andnode -erequireunsafe_shell_commands=true preview_urlmust 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:
stdio smoke:
Built-in doctor:
Docs
Distribution State
Current public install path: