An assistant that only remembers the chat will guess. A server it can call has to show its work. I built two public MCP servers on this site, and anyone can connect them to Claude, ChatGPT, Cursor or VS Code in a minute. Here is why, and how.
Why a URL, not a longer prompt
A chat window is a bad database. The model can sound sure about a role, a date, or a weather report it never looked up. MCP is the small correction. The client calls a tool. The tool returns structured data. The answer is supposed to stay inside that data, with a place to check it.
One server can serve many clients. The same HTTPS URL works from Claude, ChatGPT, Cursor, Codex, VS Code, and Gemini CLI. I keep the other client snippets on the MCP page, so this post does not become a manual.
I have shipped this shape before. In May 2026 the analyzers at NetApp gained one MCP surface and a JSON API, so Slack, the web UI, and IDE agents share one payload. I wrote about that idea in One payload, three clients. The public servers here are the same habit, pointed at pages anyone can read.
Two servers, both read-only
Ask Mario's CV is the first. A recruiter can ask their assistant about my experience and set it next to a job description. It has seven tools.
Read the CV
get_profilereturns the headline, location, skills, languages, and links.search_experiencesearches my work history and projects by keyword.list_projectslists my MCP servers, systems, gems, and experiments.get_project_detailsreturns the README, version, license, and RubyGems downloads for one project.
Compare it with a job
match_jobcites the parts of the CV that match a job description and lists what it cannot evidence. It does not return a fit score.
Posts and availability
list_postslists my blog posts with dates and links.get_availabilityreturns where I can work and how to reach me.
The data is compiled when the site is built. Streamable HTTP, no authentication, because nothing here is private.
Iceland live is the second. It has 36 read-only tools for public Icelandic data, in seven groups.
Earth and sky
get_recent_earthquakeslists recent quakes by magnitude and region.get_volcano_statusgives aviation colour codes and alert levels for each volcanic system.get_eruption_hazardssums up the Reykjanes hazard map, the Grindavík alert phase, and nearby road closures.get_aurora_forecastreturns the aurora activity index and the cloud cover for a place.get_avalanche_warningsgives avalanche danger by mountain region.
Sources: Icelandic Meteorological Office (CC BY 4.0), Almannavarnir, Grindavíkurbær, and Vegagerðin.
Weather and roads
get_weatherreturns the latest station reading and a short forecast.get_weather_warningslists active weather and hazard warnings, or says there are none.get_road_conditionsgives road surface conditions and closures, hazards first.get_highland_roadsreports on the F-roads, and says so when they are out of season.
Sources: Icelandic Meteorological Office and Vegagerðin (CC BY 4.0).
Sea and travel
get_tidesgives high and low tides for Reykjavík and 49 other harbours.get_ferriesreturns ferry schedules and reroute or cancellation notices.get_kef_flightslists Keflavík arrivals and departures with their status.get_bus_departuresreturns the next buses at a Strætó stop.
Sources: tide.is (Icelandic Coast Guard tables and TICON-4, CC BY 4.0), the ferry operators, Isavia, and Strætó open data.
Outdoors
get_ski_areasreports whether Bláfjöll, Skálafell, Hlíðarfjall, and Tindastóll are open, with lifts and on-hill weather.get_hiking_trailsfinds hiking trails, with length and ascent when the source has them.get_hot_springslists geothermal lagoons, baths, and natural hot springs, with hours and website.get_water_temperaturegives the sea temperature at Nauthólsvík beach and the published pool and hot tub temperatures.
Sources: the ski areas, Umhverfisstofnun and Landmælingar Íslands (CC BY 4.0), © OpenStreetMap contributors (ODbL), Nauthólsvík, and City of Reykjavík.
City life
get_eventslists events, concerts, exhibitions, and nightlife such as open mic, comedy, and pub quiz.get_cinema_showtimesreturns what is on in Reykjavík cinemas, with a booking link per screening.get_restaurantslists restaurants with cuisine, opening hours, and whether they are open now. It copies no reviews.get_attractionsgives museum and landmark hours, prices, and an official ticket link where one exists.get_librariesreturns Reykjavík City Library branch hours and upcoming library events.get_poolsgives swimming pool opening hours and whether a pool is open now.get_newsreturns headlines with a link, never the article text.get_air_qualityreturns hourly readings, including volcanic gas.
Sources: daily listings stored in Workers KV and marked stale after 36 hours; © OpenStreetMap contributors (ODbL); museum and Harpa websites; Borgarbókasafn; RÚV, mbl.is, and Grapevine RSS; City of Reykjavík; Environment and Energy Agency of Iceland.
Practical
get_fuel_pricesgives petrol and diesel prices per station, and the cheapest one nearby.get_ev_chargersfinds EV chargers with sockets and power. Live occupancy is not published openly, and the tool says so.get_parkingreturns Reykjavík street parking zones P1 to P4, with prices, paid hours, and city garages.get_shop_hoursgives supermarket hours, with a warning on public holidays and half-days.get_pharmacieslists pharmacies nearby, the 1700 health line, and after-hours clinic hours.get_safety_inforeturns the emergency numbers, Civil Protection notices, and active weather alerts.get_city_noticeslists water, heating, and power outages, street closures, and city notices in Reykjavík.get_waste_collectiongives the next bin collection for a Reykjavík address and recycling centre hours.get_isk_ratesgives the Central Bank of Iceland reference rates for the króna.
Sources: Gasvaktin (MIT); © OpenStreetMap contributors (ODbL); Reykjavík Parking Service; Læknavaktin, 112.is, and Almannavarnir; Icelandic Meteorological Office; Veitur; City of Reykjavík; SORPA; Central Bank of Iceland.
Calendar and daylight
get_public_holidayslists public holidays, the half-days on 24 and 31 December, and flag days.get_daylightgives sunrise, sunset, solar noon, and twilight for a town and date.
Both are calculated in code: the holidays from Icelandic law, the daylight with a standard solar algorithm.
Each payload carries a source, a fetched_at_utc timestamp, and a stale flag when the cache or the upstream is old. Booking links are always https. The server is listed in the official MCP registry as net.marioruiz/iceland.
The failure mode is easy to picture. Without a tool, an assistant can tell you a road is open from memory alone. get_road_conditions returns the official status, with a source and a fetched_at_utc timestamp. It does not invent a condition.
One URL per server. Many clients. No second copy of the facts.
Plug one in
Cursor takes a few lines in .cursor/mcp.json. The same URLs work for the other clients. The MCP page has those snippets.
{
"mcpServers": {
"marioruiz-cv": {
"url": "https://marioruiz.net/mcp/cv"
},
"marioruiz-iceland": {
"url": "https://marioruiz.net/mcp/iceland"
}
}
}
Two prompts from the server pages:
- Search Mario's CV for hands-on MCP (Model Context Protocol) experience and cite matching roles. That one calls
search_experience. - What's the weather in Reykjavík right now: temperature, wind, and precipitation? That one calls
get_weather.
The Iceland numbers move, so the example on the server page shows when it was captured. Ask again and you get the current reading.
What I want to keep doing
These are habits from the code, not a manifesto:
- The documentation and the endpoint share a URL. A GET asking for text/html gets the page; a POST speaking JSON-RPC gets the protocol. I got tired of a docs site and an API that drift apart.
- The servers are read-only and deterministic. There is no model inside them writing a row. The tools are marked read-only. The CV server's instructions say not to invent scores, dates, or skills that the tools did not return.
- Provenance and limits travel with the answer. CV payloads include their sources. Iceland payloads include the provider, the fetch time, and a stale warning. A request is capped at 32 KiB. A tool returns at most 20 items. Each client IP gets 60 requests a minute.
- Each tool has tests. If a field changes shape, I want a failing test, not a confident paragraph.
What is next
More servers, with data that is useful and public. The next one covers open scientific measurements of microplastics in the North Atlantic, and a specialist is reviewing the sources.
Which fact in your own work should an assistant have to look up, instead of remembering?