MCP servers
Optional tool servers, set up through one manager.
Chapter 1
Never in a default install
Say "set up MCP servers" after installing, and the manager asks what you want.
Why they are opt-in
Every server adds an external dependency, often a key, and its tool schemas cost tokens.
A missing or failing server falls back to the normal flow. It never blocks a task.
Free here means a local model or a free tier. Hosted services have limits or need billing.
Chapter 2
One manager, one set of rules
Each server is a data file. The shared rules live in the mcp-manager skill.
The setup interview
The manager asks one question at a time and skips any you already answered: which servers, which AI tools, which scope, which keys, and how to route between a server and the normal flow.
The routing rule is MCP first, then fallback. A server file lists what it replaces and when it is used first.
Server 1 of 4
image-gen
Generates images for sites and apps.
Does
Image generation for websites and apps. It also has enhance_image, which grades, levels and sharpens an existing photo with ImageMagick and no AI, and ai_edit_image for small AI edits.
You
Set at least one set of keys: GEMINI_API_KEY (needs billing), or CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN, which have a free allowance of about 173 images a day.
AI
Tries Gemini, then Cloudflare FLUX.1 schnell, then a labelled placeholder. It tells you about every switch and pause.
Server 2 of 4
local-vision
Reads images offline for models that cannot see them.
Does
Free offline image reading with Qwen3-VL through llama.cpp, as a stdio tool. It gives models without vision a way to read screenshots, diagrams and photos.
You
Run node local-vision.js setup. It downloads about 1.6 GB for the 2B model, or about 3 GB for 4B on GPUs with 6 GB or more. A GPU is optional.
AI
Used first when the model lacks vision, and only on request for Claude, which already reads images. Without it, the image is skipped or you are asked to describe it.
Server 3 of 4
stitch
UI prototyping with Google Stitch.
Does
AI UI design generation, with screens, variants and design systems, for rapid prototyping before code.
You
A remote HTTP server. Put an API key from stitch.withgoogle.com/settings in STITCH_API_KEY. It is free for now with usage limits, and that may change.
AI
Tried first for new mockups and layout exploration. The fallback is a Claude artifact or a plain HTML prototype.
Server 4 of 4
prefilter
Ranks candidate files before the agent reads them.
Does
Exposes the System-1 prefilter add-on as a tool, so the agent can rank candidates before it opens them.
You
Local and free with the Laya model. It starts on first use and is set up through the add-on's own setup.
AI
If it is not configured, the agent skips prefiltering and does the task the normal way.
Finale
Not sure which one?
Start from the situation.
You need images for a site or app
image-genThe model cannot see images
local-visionYou want UI mockups before any code
stitchYou want fewer file reads before work starts
prefilter
Keep reading
On to the next page
The docs are a short run of pages. Use the buttons, or jump to any page.