touch-grass-mcp

Preference-aware events discovery MCP server. Surfaces what you'd actually want to do — not what's loudest.

touch-grass-mcp is a Model Context Protocol server that aggregates events, restaurants, museums, music, and trending venues across multiple sources, then re-ranks results against your taste profile so you stop scrolling generic event listings and start finding things you'd actually go to.

Self-hosted by design. Zero telemetry. All data stays on your machine.

What you get

Core (works in any US city):

• Concerts (Ticketmaster, Resident Advisor, Dice)

• Restaurants + bars (Yelp)

• Local groups + meetups (Meetup, Eventbrite)

• Theater (TodayTix)

• Breweries (Open Brewery DB)

• Weather (NWS)

• Editorial picks (Eater, Gothamist, Time Out, Hyperallergic, Artforum)

NYC pack (activates when your profile city is New York):

• Public libraries (NYPL, BPL, QPL)

• Museums (MoMA, Met, Whitney, Frick, Carnegie Hall, 92Y, Park Avenue Armory, MoMA PS1)

• Venues (Lincoln Center, Brooklyn Steel, Metrograph, Village Vanguard, Village Jazz)

• NYC Open Data (city-permitted events)

• Audubon (birding walks)

• Curated community calendars

The taste engine — what makes this different from a flat aggregator:

• Profile-driven scoring: your interests, dislikes, vibe preferences, neighborhoods

• City pulse: live trend tracker (Reddit + Google Trends + RSS) that surfaces what's rising

Install

pip install touch-grass-mcp
# or with optional pulse + NLP support:
pip install "touch-grass-mcp[pulse,nlp]"

First-run setup

# One unified onboarding wizard: profile + city pack + API keys
touch-grass setup

# Fast launch for a known city (auto-fills state + pulse defaults):
touch-grass setup --city "san francisco"

# See available city starter packs:
touch-grass list-cities

# Already configured profile? Just (re)do API keys:
touch-grass setup --keys-only

# Sanity check:
touch-grass doctor

City starter packs auto-fill subreddits, RSS feeds, and Google Trends geo when your profile city matches. Currently bundled: NYC (deep coverage), SF, LA, Chicago, Austin, Boston, Seattle, DC. PRs welcome to add more.

Wire it into Claude Desktop / Claude Code

Add to your MCP config (~/Library/Application Support/Claude/claude_desktop_config.json or ~/.claude.json):

{
  "mcpServers": {
    "touch-grass": {
      "command": "touch-grass",
      "args": ["serve"]
    }
  }
}

Tools exposed

23 MCP tools across 5 categories:

Event search: search_events, search_concerts, discover_niche_events, trending_events, get_event_details, search_community_calendars, search_ra_events, get_ra_event_details

Dining + venues: search_restaurants, search_breweries, get_restaurant_details

Arts + culture: search_broadway_shows, get_broadway_showtimes, get_museum_exhibitions, search_met_collection, get_editorial_picks, get_editorial_feed

Profile + recommendation: get_user_profile, update_user_preferences, get_recommendation_keywords

Calibration + utility: weekend_weather, log_flag_feedback, get_calibration_stats

Configuration

• Profile: ~/.config/touch-grass/config.json (run touch-grass init to bootstrap)

• API keys: ~/.config/touch-grass/.env (or shell env)

• Cache + state: ~/.local/share/touch-grass/ (XDG paths)

See CONFIG.md for the full profile schema.

Privacy

Read PRIVACY.md before installing. Short version: zero telemetry, no analytics, no phone-home, all data local.

Architecture

• taste/ — pure-Python preference engine: load profile, score events, rank by relevance. Zero MCP dependencies; importable from any context.

• pulse/ — cultural trend tracker that re-ranks against current Reddit / Trends / RSS signal.

• packs.py — city-pack registry. NYC pack bundled. Adding a new city is a documented contribution.

• server.py — FastMCP server that wires it all together.

Cities other than NYC

The core pack works anywhere in the US. NYC is the bundled deep-coverage city. Adding a new city = drop scrapers in clients/<city>/, register a CityPack in packs.py, supply pulse defaults. PRs welcome.

License

MIT.

Status

v0.1 — alpha. Stable enough for personal use; expect breakage in fragile scrapers (museums, RSS).