The CLI is also an MCP server. Connected, a coding agent can search the catalogue, read every control a shader exposes — range, default, and what moving it does — and vendor tuned source straight into the project it is working on. It is npx shadertown add with the judgement moved into the conversation.

Connect an agent

npx add-mcp "npx -y shadertown@latest mcp" -n shadertown

add-mcp detects the coding agents on the machine — Claude Code, Cursor, VS Code, Windsurf, Codex, Zed, and some twenty others — and writes each one's configuration in its own dialect. Pass -a cursor (repeatable) to target one explicitly, or -g for the user-level config instead of the project's. Any MCP client takes the same server by hand:

{
  "mcpServers": {
    "shadertown": {
      "command": "npx",
      "args": ["-y", "shadertown@latest", "mcp"]
    }
  }
}

The server speaks MCP over stdio and adds no dependencies — it is the same zero-dependency CLI, answering discovery from the registry the package already ships and installing through the same registry route as add.

Or connect to the hosted one

There is a second server, at https://www.shadertown.com/mcp, speaking the same protocol over Streamable HTTP. Nothing is installed and nothing runs on the machine:

{
  "mcpServers": {
    "shadertown": {
      "url": "https://www.shadertown.com/mcp"
    }
  }
}

Which one to reach for is a question of what the agent is allowed to do. The hosted server can only hand back bytes — it has no access to the project, so it answers with the files' contents and the agent writes them. The stdio server runs where the project is, so it writes them itself. Same catalogue, same paywall, same licence:

	`npx shadertown mcp`	`https://www.shadertown.com/mcp`
Transport	stdio	Streamable HTTP
Install	none, via `npx`	none
Reads the catalogue	yes	yes
Writes files into the project	`install_shader`	returns their contents
Signs in	the machine's stored token	an `Authorization` header

The hosted server takes a device token in an Authorization: Bearer header for Professional source. Without one it serves the Free shaders and explains what the others need.

The tools

Tool	What it does
`search_shaders`	Free-text search over names, taglines, descriptions, families, and tags
`get_shader`	One shader in full: every control with range, default and meaning, the colours, usage per framework, and the URL where a human can tune it live
`install_shader`	Vendors the WGSL, a typed preset pinned to the tuned params, a working example, the README, and the LICENSE into the project
`account_status`	Who this machine is signed in as, on which plan, and what that unlocks

The hosted server has the first two, and get_shader_source in place of the last two: it returns the same five files for the agent to write, and has no machine to report the sign-in state of.

install_shader writes the same five files as npx shadertown add, relative to the workspace the agent runs in — src/shaders by default, or the outDir from shadertown.config.ts.

Licensing

The 16 Free shaders install for anyone. The other 61 need a plan on the signed-in account — until then they answer 402:

npx shadertown login

The agent cannot run that for you — it opens a browser — but when an install hits the paywall, the tool result says exactly this, so the agent will ask. For CI or containers, set SHADERTOWN_TOKEN instead.

Things to say to it

"Put a calm, dark animated background behind the hero — pick one from shadertown and install it."
"Make the aurora slower and shift it toward the brand purple, then pin those values into the preset."
"Which shadertown shaders react to the pointer? Wire one up behind the signup form."
"Swap the vendored shader for one with a lighter per-frame cost."

MCP server

Connect an agent

Or connect to the hosted one

The tools

Licensing

Things to say to it

On this page