espresso-mcp

An MCP server that finds great espresso cafes — and codifies what makes them great.

espresso-mcp is a Model Context Protocol server you can plug into Claude Desktop, Claude.ai, ChatGPT, Google Gemini, Cursor, Hermes, OpenClaw, and any other standard MCP host. It ships with a curated database of 100+ specialty-coffee shops, 75+ specialty roasters, and a transparent scoring algorithm that captures why a shop is good — sourcing, freshness, equipment, training — and why a shop is bad (the heaviest negative signal: menus dominated by flavored syrups).

It is not a Yelp clone. The data is hand-curated, the scoring is explicit, and the philosophy is "if it's covering bad coffee with flavors, we don't want it."

---

The bias

espresso-mcp has an opinion. Shops that hide bad coffee behind flavored syrups are the failure pattern this tool is designed to filter out. The scoring algorithm gives syrup_emphasis the heaviest non-disqualifying penalty (−22) — heavier than any single positive signal. Mass-market chains (Starbucks, Dunkin', Costa, Tim Hortons, Peet's, Caribou) and "looks third-wave but tastes flavored" shops (the algorithm's flavor-led-specialty category) score in avoid or fair tier regardless of how good the signage looks.

Positive credit goes to the signals great shops actually invest in:

• In-house roasting, or named-partner sourcing from a known specialty roaster
• Single-origin espresso (not just blends)
• Roast date on retail bags (peak-freshness commitment)
• Espresso + pour-over both offered — the "complete program"
• Cortado on the menu — a confidence proxy; shops avoid small milk drinks when their espresso can't hold up
• Competition involvement / SCA-certified staff
• Equipment from the third-wave shortlist (Slayer, La Marzocco, Synesso, Decent, Victoria Arduino, Modbar)
• Recognition from authoritative rankings (World's 100 Best, SCA championships, Coffee Review 95+, Good Food Awards)

Every signal is documented at src/scoring/weights.ts and surfaced to clients via the score breakdown on every result — you can see exactly why a cafe is recommended.

---

See it in action

You: I'm near Brandenburg Gate in Berlin. Find me 3 great espresso cafes 
     within walking distance.

Claude: [resolves Brandenburg Gate → 52.5163, 13.3777]
        [calls find_espresso_near with radius_km=2]

  Found 3 specialty cafes within 2km of Brandenburg Gate, sorted by quality:

  • The Barn (Mitte)         — 0.8km — score 95 (world-class) — World's #61
    In-house roastery, single-origin focus, "never blends beans." Founded 
    2010, one of Europe's leading third-wave roasters.

  • Companion Coffee (Mitte) — 1.1km — score 83 (great)
    Curated multi-roaster with rotating guest beans. Tea + coffee crossover 
    in a quiet Mitte design space.

  • Westberlin (Mitte)       — 1.3km — score 82 (great)
    Iconic third-wave anchor since ~2010. Magazine + coffee concept; pours 
    The Barn and rotating European roasters.

---

What it does

Once installed in your MCP client, you can ask things like:

• "I'm at the Hotel Adlon Berlin — find me great espresso within walking distance."
• "I'm staying near the Brandenburg Gate. Recommend 3 cafes and rank by quality."
• "Search for cafes in Tokyo scoring above 80."
• "Tell me about Tim Wendelboe."
• "Score this cafe — they have a Slayer, in-house roasting, single-origin espresso, and no flavored syrups."
• "What are the world-class roasters in Denmark?"
• "What should I avoid in Chicago? Show me an anti-pattern example."

The model gets a structured score with reasoning, distance, awards, and per-signal contributions — enough to give you an honest recommendation rather than a popularity list.

---

Quick install

Client	Section
Claude Desktop	↓
Claude.ai (Browser MCP)	↓
ChatGPT	↓
Google Gemini CLI	↓
Hermes (Nous Research)	↓
OpenClaw	↓
Cursor	↓
VS Code	↓
Any standard stdio	↓
From source	↓

Recommendation: use @latest while the project is iterating

While we're still adding cafes and refining the algorithm, pin to the live npm tip:

"args": ["-y", "espresso-mcp@latest"]

Once the data and scoring stabilize, you can drop @latest and pin a specific version for reproducibility.

---

Per-client configuration

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows). If the file doesn't exist, create it:

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

Restart Claude Desktop. The 🔌 menu should show 6 tools under "espresso."

Claude.ai (Browser MCP)

Claude.ai's web app supports MCP servers via OAuth and remote endpoints. For local installation, use Claude Desktop instead (above). To expose espresso-mcp to Claude.ai as a remote server, wrap it with mcp-remote (see the ChatGPT section below — same approach).

ChatGPT

⚠ Important: ChatGPT only supports remote MCP servers (HTTPS endpoints), not local stdio processes. You have two options:

Option 1: Use the hosted version (when available). If we publish a hosted endpoint we'll list it here.

Option 2: Bridge espresso-mcp to HTTPS via mcp-remote. Run a small bridge on a machine you control:

npx -y mcp-remote bridge espresso-mcp \
  --port 8080 \
  --token "your-shared-secret"

Then expose port 8080 via a tunnel (Cloudflare Tunnel, Tailscale Funnel, or a small VPS) and use the resulting HTTPS URL.

Enabling MCP in ChatGPT (Plus / Pro / Team / Enterprise plans only):

1. Settings → Connectors → Advanced → toggle Developer mode on.
2. Add a custom connector pointing at your bridge URL.

OpenAI's MCP docs have the latest connector setup.

Google Gemini CLI

Edit ~/.gemini/settings.json:

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

Restart Gemini CLI. It will auto-connect at startup and show "Connected" if successful. See Gemini CLI MCP docs for the full reference.

Hermes (Nous Research)

Hermes uses YAML, not JSON. Edit ~/.hermes/config.yaml:

mcp_servers:
  espresso:
    command: "npx"
    args: ["-y", "espresso-mcp@latest"]

Optional: filter to just the tools you want exposed:

mcp_servers:
  espresso:
    command: "npx"
    args: ["-y", "espresso-mcp@latest"]
    tools:
      include: [find_espresso_near, search_cafes, score_cafe]

Restart Hermes — it auto-discovers MCP tools at startup. See Hermes MCP docs.

OpenClaw

OpenClaw uses standard stdio MCP config. Install OpenClaw if you haven't:

curl -fsSL https://openclaw.ai/install.sh | bash
# or: npm install -g openclaw@latest

Then add to your OpenClaw config (typically ~/.openclaw/config.json — check OpenClaw MCP docs for the current location):

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

Cursor

Edit ~/.cursor/mcp.json:

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

VS Code

VS Code uses a dedicated mcp.json file (not settings.json). For workspace-scoped config, create .vscode/mcp.json. For user-wide, run the MCP: Open User Configuration command from the Command Palette.

{
  "servers": {
    "espresso": {
      "command": "npx",
      "args": ["-y", "espresso-mcp@latest"]
    }
  }
}

Note: VS Code uses the key servers (not mcpServers) — this is the most common copy-paste mistake when migrating configs from other clients.

Generic stdio client

For any MCP-compatible host that spawns local processes:

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

The server writes JSON-RPC to stdout and logs to stderr only — capture stderr if you want startup messages.

From source (development)

git clone https://github.com/mattgierhart/espresso-mcp
cd espresso-mcp
npm ci
npm run build

# Point your client at:
#   command: node
#   args:    ["/absolute/path/to/espresso-mcp/dist/index.js"]

Or run the MCP Inspector against the local build:

npm run inspect

---

Travel & address-based queries

The most common real-world question is "I'm at X, where should I get coffee?" Here's the pattern in practice.

Hotel-based recommendation

You: I'm staying at the Hotel Adlon Berlin. Find me 3 great espresso cafes nearby.

[Claude searches for the Adlon's coordinates: ~52.5163, 13.3795]
[Claude calls find_espresso_near with those lat/lon]
[Returns ranked Berlin cafes within walking distance]

The model resolves the address to coordinates (using its built-in knowledge or a web search), then calls find_espresso_near. No extra config needed.

Address-based query

You: I'm walking around Millennium Park in Chicago. What's the closest specialty espresso?

[Claude resolves the landmark → ~41.882, -87.622]
[Calls find_espresso_near with radius_km: 2]
[Ranks results by quality, then distance]

Multi-city trip planning

You: I have trips coming up to Tokyo, Berlin, and Hong Kong. Build me a coffee plan —
     top 2 cafes per city, plus the one roaster I should buy beans from in each.

[Claude calls search_cafes for each city, sorts by score]
[Calls list_great_roasters filtered by country]
[Composes a per-city itinerary]

Score what you're looking at right now

You: I'm at a cafe with a Slayer machine, they roast on-site, single-origin espresso 
     option, roast date on bags, and they only have plain milk drinks — no syrups. 
     Score it.

[Claude calls score_cafe with those signals]
[Returns 95/100 ("world-class") with per-signal contributions]

Hotel coffee-walkability comparison (manual, until v0.4)

You can do this today with multiple tool calls:

You: Compare these Berlin hotels by walkable specialty coffee:
     - Hotel Adlon (Mitte)
     - Soho House Berlin (Mitte)
     - 25hours Hotel Bikini (Charlottenburg)

[Claude calls find_espresso_near three times, one per hotel]
[Composes a comparison table]

A dedicated score_hotel_coffee_access tool is on the roadmap for v0.5.

Best-effort with cities not yet in the database

If you query a city we haven't curated yet (e.g., Lisbon, Seoul, Mexico City), find_espresso_near returns no curated matches. The model can still combine its own knowledge with score_cafe to evaluate any cafe you describe. We're filling in cities one batch at a time — open an issue if you want yours prioritized.

---

Tools

find_espresso_near

Find ranked specialty espresso cafes within a radius of coordinates.

{ "lat": 35.6855, "lon": 139.6904, "radius_km": 3, "min_score": 60, "limit": 10 }

Returns cafes from the curated database sorted by espresso-quality score, with per-cafe distance and reasoning.

search_cafes

Search the curated database by query, city, country, roaster, or minimum score.

{ "query": "natural wine", "country": "DE", "min_score": 70, "limit": 20 }

get_cafe_details

Full record for a cafe by id, including the score breakdown signal-by-signal and a few related/nearby cafes.

{ "id": "tim-wendelboe-oslo" }

score_cafe

The codified algorithm exposed directly. Pass in signals you've observed (from a website, a review, a photo) and get a 0-100 score with a per-signal contribution breakdown. No database lookup required.

{
  "name": "Hypothetical Shop",
  "observed_signals": {
    "roasting": "in-house",
    "brew_methods": ["espresso", "pour-over"],
    "single_origin_espresso": true,
    "roast_date_on_bags": true,
    "cortado_on_menu": true,
    "syrup_emphasis": false
  }
}

list_great_roasters

Curated specialty roasters by country and reputation tier.

{ "country": "DK", "min_reputation": "regional-leader", "limit": 25 }

list_anti_patterns

Curated shops that exemplify what to avoid — the contrast set for the algorithm. Two flavors:

• mass-market-chain — Starbucks, Dunkin', Costa, Tim Hortons, Peet's, Caribou. Generic dark roasts, flavored-drink menus, low sourcing transparency.
• flavor-led-specialty — shops that display third-wave signage (in-house roasting, single-origin signs, roast dates) but in practice serve a syrup-forward menu. Looks specialty, drinks flavored. Useful contrast when explaining why a recommended shop is the real thing.

{ "category": "flavor-led-specialty", "limit": 10 }

Anti-patterns are stored separately in data/anti-patterns.json so they never bleed into find_espresso_near or search_cafes results.

---

How scoring works

The full weight table is the source of truth at src/scoring/weights.ts. In short:

Positive signals (the things great shops do):

• In-house or named-partner roasting from a known specialty roaster
• Single-origin espresso
• Roast date on retail bags (peak-freshness commitment)
• Espresso + pour-over both offered ("complete program")
• Cortado on the menu (confidence in espresso)
• Competition involvement / SCA-certified staff
• Quality espresso machines (Slayer, La Marzocco, Synesso, Decent, Victoria Arduino, Modbar)
• Awards (World's 100 Best, SCA championships, Coffee Review 95+, Good Food Awards)

Negative signals (the cover-up patterns):

• syrup_emphasis (−22) — menu dominated by flavored-syrup drinks. The strongest avoid-signal. Great shops don't hide behind syrups.
• flavored_drink_share > 0.5 (additional −7)
• no_grinder_visible (−25) — effectively disqualifying
• only_dark_roast (−8) — masks bean defects
• no_origin_info (−7) — they don't know or don't care

Roaster reputation bonus: cafes that roast in-house or partner with a known roaster get +25 (world-class), +15 (regional-leader), or +8 (notable) on top of structural signals.

Null/unknown signals are skipped, not penalized. The score includes a confidence value proportional to how many signals were actually observed.

Score tiers:

Score	Tier
≥ 85	World-class
≥ 70	Great
≥ 55	Good
≥ 40	Fair
< 40	Avoid

The "looks specialty but isn't" pattern is tricky — a cafe with genuine third-wave structural signals plus a flavored-drink menu will score in the "good" range despite the syrup penalty. For those cases, the explicit category: flavor-led-specialty flag in data/anti-patterns.json is the authoritative human override.

---

Curated coverage (v0.2.0)

The database is hand-curated. Coverage skews toward cities where we (or trusted curated guides) have actual recent visits. We add cities batch-by-batch — open an issue if you want yours prioritized.

Region	Cafes	Districts represented
Berlin	15	Mitte, Kreuzberg, Neukölln, Charlottenburg
London	15	Holborn, Shoreditch, City, Marylebone, Bishopsgate, Hammersmith (incl. Workshop × 3)
NYC (Manhattan + Brooklyn)	17	East Village, Nolita, Noho, Flatiron, FiDi, NoMad, Williamsburg, Bushwick, Park Slope
Dallas / DFW	18	Design District, Lower Greenville, Oak Cliff / Bishop Arts, Deep Ellum, Knox-Henderson, Plano
Denver metro	18	Capitol Hill, RiNo, LoHi, Tennyson, Edgewater, Aurora, Lone Tree, Littleton
Hong Kong	10	Causeway Bay, Sheung Wan, Tsim Sha Tsui, Tai Hang
Tokyo / Osaka	6	Omotesando, Nakameguro, Kuramae, Shimokitazawa, Sangubashi
Guangzhou	8	Tianhe, Yuexiu, Liwan
Other	~9	Oslo, Copenhagen, Dublin, Galway, Vienna, Singapore, Santa Ana SV, Rogers AR

Plus 8 deliberately-flagged anti-pattern entries (Starbucks, Dunkin', Costa, Tim Hortons, Peet's, Caribou, plus two "looks specialty but tastes flavored" examples).

Gaps worth filling next: SF, LA, Seattle, Portland, Melbourne, Sydney, Stockholm, Amsterdam, Paris, Taipei, Seoul, Mexico City.

---

Contributing a cafe

Open a PR against data/cafes.json:

1. Add the entry in alphabetical order by city, then name.
2. Use exact coordinates when possible ("coord_precision": "exact"). Look them up on Google Maps.
3. Set last_verified to today's ISO date.
4. Only include signals you've personally observed or can cite.
5. If you're naming a new roaster, add it to data/roasters.json with a reasonable reputation tier (notable, regional-leader, or world-class).
6. Run npm test and npm run validate-data before opening the PR.

See data/README.md for the full schema reference.

Suggesting an anti-pattern

Anti-patterns belong in data/anti-patterns.json with a category field set. Strong candidates are shops you've personally verified as either mass-market or "looks specialty but drinks flavored." Include a notes field explaining the gap between signage and experience.

Reporting feedback

Open an issue with what you tested, what worked, what surprised you, and what's missing. We track the roadmap (community contribution tool, automated address geocoding, live data fallback, hotel-proximity scoring) as open issues.

---

License

• Code is MIT — see LICENSE.
• Data (data/*.json) is CC-BY 4.0 — see data/LICENSE. Attribution required.

If you redistribute the data, please link back to this repository.

---

Sources we trained the patterns on

The curated cafe roster, the signal weight table, and the anti-pattern category framing were built by reading and cross-referencing the people who already do this work well. Credit and links:

City and region guides

• World's 100 Best Coffee Shops — annual industry-voted ranking; the backbone of the world-class tier
• European Coffee Trip — 6,229+ curated European specialty cafes; used heavily for the London and Berlin batches
• D Magazine — Best Dallas Coffee Shops and Roasters — the canonical DFW specialty guide
• Drips of God — New York — Manhattan + Brooklyn curation
• Bee An Coffee — Best Specialty Coffee in NYC 2026 — Brooklyn roastery focus
• Grounds Club — Hong Kong Specialty Coffee Guide — HK-specific
• Asian Coffee Map — Asia regional anchor

Awards and rankings

• Sprudgie Awards — Sprudge's annual community awards
• Good Food Awards: Coffee — US specialty roaster recognition
• Coffee Review (95+ scores) — exceptional roast quality
• SCA World Barista Championship — competition involvement is a high-signal indicator

Frameworks and taxonomies

• Sprudge — long-form specialty coffee journalism; the original "signal taxonomy" intuition came from years of Sprudge reading
• Specialty Coffee Association — Barista Skills curriculum — equipment shortlist and certification framework
• SCA Certified Commercial Equipment list — what "quality espresso machine" means

A lot of time spent in cafes. Fieldwork. Probably not tax-deductible. The most honest entry on this list — every weight in src/scoring/weights.ts was calibrated by someone who has been personally disappointed by a Slayer-equipped shop that turned out to lean syrup-forward.

Built on

• @modelcontextprotocol/sdk — the official MCP TypeScript SDK
• Zod, tsup, Vitest, Firecrawl (for scraping city guides during curation)