mcpField
P
pls-touch-grass

pls-touch-grass

Local
github.com/RinavaGames & Entertainment0Updated 10w ago

About pls-touch-grass

pls-touch-grass is a humorous MCP server that tracks outdoor breaks and nudges people to leave their screens. It is also a demonstration of MCP tool calls and shared state, combining frustration detection with weather and nearby park lookups.

Key features

  1. 01Log outdoor visits
  2. 02Track time spent indoors
  3. 03Detect frustration and escalate
  4. 04Check weather and nearby parks
  5. 05Persist state in a JSON file

Use cases

  1. 01Log a return from the park
  2. 02Check whether it is time to go out
  3. 03Trigger a break reminder during frustration

MCP support

It provides five tools to MCP clients over stdio for logging outdoor visits, checking indoor time, detecting frustration, checking weather, and finding parks.

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

mcp.json · Universal config
{
"mcpServers": {
"rinava-pls-touch-grass": {
"command": "npx",
"args": [
"-y",
"pls-touch-grass-mcp"
],
"env": {
"GRASS_STATE_FILE": "<GRASS_STATE_FILE>"
}
}
}
}
Run commandnpx -y pls-touch-grass-mcp
Environment variables to set
GRASS_STATE_FILEPath to the JSON file holding your grass-touching history. Defaults to ~/.pls-touch-grass.json

Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.

Packages
npmpls-touch-grass-mcp0.3.0

Tools

5

Saved tool list · updated 2w ago

ParameterTypeRequiredDescription
notestringNoWhat they did outside, e.g. 'went to the park'

Details

From the project README

pls-touch-grass-mcp

npm ci license: MIT

An MCP server that knows how long it's been since you touched grass, and won't shut up about it.

You log grass, it tracks indoor time, it catches you raging at your AI and sends you to the nearest park, weather checked first. A joke, but one that teaches MCP: five tools, shared state, and descriptions that make the model call them unprompted.

Install

Claude Code
claude mcp add pls-touch-grass -s user -- npx -y pls-touch-grass-mcp

-s user, or the server only exists in the directory you ran that from. Working from a clone, skip the command: there's a .mcp.json in the root and Claude Code offers it when you open the folder.

Cursor

Add to Cursor

Or the JSON, in ~/.cursor/mcp.json (global) or .cursor/mcp.json (per project):

{
  "mcpServers": {
    "pls-touch-grass": {
      "command": "npx",
      "args": ["-y", "pls-touch-grass-mcp"]
    }
  }
}
Claude Desktop

.mcpb from Releases, then Settings → Extensions → Install Extension. Dependencies are bundled; installs offline. By hand it's the same JSON as Cursor, in claude_desktop_config.json (~/Library/Application Support/Claude/ on macOS, %APPDATA%\Claude\ on Windows).

claude.ai

No. Custom connectors want a remote server over HTTPS; this is stdio on your own machine. HTTP transport is the first item under Homework.

From source
git clone https://github.com/Rinava/pls-touch-grass-mcp
cd pls-touch-grass-mcp
npm ci && npm run build

Point any config above at node /absolute/path/to/dist/index.js instead.

npx -y downloads the package on first start — live audiences and venue wifi deserve a warm cache. Published on npm as pls-touch-grass-mcp, in the MCP Registry as io.github.Rinava/pls-touch-grass.

The five tools

ToolWhat it doesWhen the model calls it
touched_grassLogs that you went outside"back from the park"
gotta_goDo you need to go out?"can I keep going?"
frustration_detectorCatches the rage and escalateson its own, when you snap; that's the point
grass_conditionsReal weather (Open-Meteo, no key)before sending you outside
where_to_touch_grassNearby parks with distances (OpenStreetMap)"where should I go?", or when you name a place

Nobody teaches the model to detect frustration: the tool description does it alone. Type "NOTHING WORKS!!!" and watch.

Zero configuration

Location resolves fresh from your IP (ipwho.is) each run — city-level on purpose. Detection down: last known spot, then the Obelisco, like everyone else. Name a neighborhood or a street address to where_to_touch_grass and it sticks: known neighborhoods resolve offline, anything else is geocoded once through OpenStreetMap's Nominatim, and only the words you typed leave your machine. A place you named outranks a guess from your IP, until you name a different one.

Indoor tolerance defaults to a merciless 2 minutes; ask the model for clemency and it passes a bigger threshold_minutes to gotta_go. State is one JSON file, ~/.pls-touch-grass.json. Touching grass resets the rage counter. That's how absolution works.

The demo curse

Live demos summon it: the venue wifi dies the moment you say "watch this". Add --demo to the args and every answer goes canned — 22°C and clear, the Obelisco, the curated spots. Zero network calls, zero surprises. Off by default; real life should stay real.

Follow the workshop

The repo is tagged by step:

git checkout step-1   # stdio server + touched_grass
git checkout step-2   # + gotta_go (state and threshold)
git checkout step-3   # + frustration_detector (the demo)
git checkout step-4   # + grass_conditions (network with fallback)
git checkout step-5   # + where_to_touch_grass and this README
npm ci && npm test

Every step compiles and passes its own tests. Two dependencies total: the official MCP SDK and zod.

Homework

  • HTTP transport: createMcpHandler in SDK v2 serves the same factory over fetch, and it's what gets you onto claude.ai
  • Per-project thresholds instead of a global one
  • Streak tracking for consecutive grass days
  • Demo spots for your own city, via PR: one object per park in src/lib/demo.ts, with an opinion

Release

npm run build:mcpb packs dist and the production dependencies into pls-touch-grass-mcp-<version>.mcpb, the file that goes on a GitHub Release. The manifest is generated from package.json; nothing to keep in sync by hand.

Contributing

Issues and PRs welcome — the Homework list is seeded as open issues, several tagged good first issue. See CONTRIBUTING.md for setup and the house rules (there are few, but the demo-spot opinion is mandatory).

License

MIT. The grass is free and so is this.

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.