Docs / Catalog

Tools, resources & prompts

What every endpoint serves, in every revision: 22 tools, four files and three URI templates, and three prompts — all deterministic, all drawn from the same fake world, each with a documented way to fail.

Determinism

tools/list returns the 22 tools in the same order on every revision. Real computations compute; fake-world tools derive their data from a hash of the normalized arguments.

The same call always returns the same bytes, on Bun and on Node, on any machine. Different calls return different but internally consistent data: metric and imperial agree, currency conversions round-trip, a 3-day forecast is the start of the 7-day one, and weather://Madrid reads what get_weather reports for Madrid.

Arguments are normalized before hashing and matching: " madrid " and "Madrid" are the same city, "eur" and "EUR" the same currency.

await client.callTool({ name: "get_weather", arguments: { city: "Madrid" } });
// structuredContent: { city: "Madrid", temperature: 14, temperature_unit: "°C",
//   conditions: "thunderstorm", humidity: 42, wind_speed: 55, wind_speed_unit: "km/h" }

await client.callTool({ name: "get_weather", arguments: { city: " madrid ", units: "imperial" } });
// structuredContent: { city: "Madrid", temperature: 57.2, temperature_unit: "°F",
//   conditions: "thunderstorm", humidity: 42, wind_speed: 34.2, wind_speed_unit: "mph" }
Arguments that break a tool's input schema are a tool execution error (isError: true) from 2025-11-25 on, and a JSON-RPC -32602 protocol error before it — as each revision's spec lists them. An unknown tool is -32602 everywhere.

Real computations

Eleven tools that actually compute their answer. All are read-only and idempotent.

ToolArgumentsReturns
echotextThe text prefixed with "Echo: ", as a text block
adda, b (numbers){ sum }
calculateexpression, up to 1000 chars{ result }. A safe parser for + - * / % ^ and parentheses — no eval
transform_texttext, operation: upper · lower · reverse · slug{ result }
text_statstext{ words, chars, lines }
sort_listitems (up to 1000 strings), order?: asc · desc{ items }, sorted by code point
convert_unitsvalue, from, to: length (mm cm m km in ft yd mi), mass (mg g kg oz lb), time (s min h d), temperature (c f k){ value, unit }
hash_texttext, algorithm: md5 · sha1 · sha256{ algorithm, hex }
date_difffrom, to: YYYY-MM-DD (UTC) or date-times with an offset{ days, hours, seconds }
color_swatchcolor: #rgb, #rrggbb or a name; size?: 1–64 px, default 16A real PNG, as an image block
simulate_worksteps?: 1–20, default 5; step_ms?: 0–2000, default 100{ steps, completed }, after really waiting between steps

simulate_work is the tool for everything that takes time: it sends notifications/progress on each step when the request carries a progressToken, it can be cancelled, it runs as a task for clients that declare the tasks extension, and a long step_ms lets a test exercise its client's timeout. See Progress & cancellation.

Fake world

Eleven tools that look up a world that doesn't exist, believably. Annotations tell the read-only lookups from the two that act: send_email and delete_file.

ToolArgumentsReturns
get_weathercity, units?: metric · imperial{ city, temperature, temperature_unit, conditions, humidity, wind_speed, wind_speed_unit }
get_forecastcity, days: 1–7, units?{ city, temperature_unit, days: [{ day, high, low, conditions, precipitation_chance }] }; day 1 is tomorrow
search_webquery, limit?: 1–10, default 5{ query, results: [{ title, url, snippet }] } plus a resource_link per result, all under example.com
get_stock_pricesymbol: 1–5 letters{ symbol, price, currency, change, change_percent }
convert_currencyamount, from, to: ISO 4217 codes{ amount, from, to, rate, result }; rates are consistent across pairs
get_useruser_id{ id, name, email, role, created_at }
geocodeaddress{ formatted_address, latitude, longitude }
get_flight_statusflight_number: airline code + 1–4 digits, e.g. IB3456{ flight_number, status, departure_airport, arrival_airport, scheduled_departure, terminal, gate }
read_filepath: absoluteThe file as an embedded resource, its contents following the extension: .json, .md, .csv, anything else as plain text
delete_filepath: absolute{ path, deleted }. Asks for confirmation first on 2026-07-28; nothing is really deleted
send_emailto, subject, body{ message_id, to, status }. Nothing is really sent
delete_file on 2026-07-28 answers its first call with input_required and a form elicitation — “Delete /tmp/old.log? This cannot be undone.” — and the retry completes or is cancelled according to the answer. A client that does not declare form elicitation gets -32021. Legacy revisions, which cannot ask, delete without asking, leaving the confirmation to the client as the destructiveHint annotation did in 2025. See Multi round-trip requests.

Failing on purpose

Each fake-world tool fails on one reserved value, so a test can reach every failure path without forcing anything. Failures are tool execution errors: isError: true with a message.

ToolFails withMessage
get_weather, get_forecastcity: "Atlantis"City not found: Atlantis
geocodeaddress: "Atlantis"Address not found: Atlantis
search_webquery: "error"Search failed: the search backend is unavailable
get_stock_pricesymbol: "XXXX"Unknown symbol: XXXX
convert_currencyfrom or to: "XXX"Unsupported currency: XXX
get_useruser_id: "missing"User not found: missing
get_flight_statusflight_number: "XX000"Flight not found: XX000
read_file, delete_filepath: "/missing"File not found: /missing — and any path with .. is refused with Access denied
send_emailto: "bounce@example.com"Delivery failed: the mailbox bounce@example.com does not exist
const result = await client.callTool({ name: "get_weather", arguments: { city: "Atlantis" } });
// { content: [{ type: "text", text: "City not found: Atlantis" }], isError: true }

For any other answer — a different error message, a protocol error, an HTTP failure — force it from the request. See Controlling answers.

Content types

Between them, the tools return every content type a tools/call result can carry, so a client's rendering of each one gets exercised.

ContentFrom
textEvery tool. Tools with structured content repeat it as JSON text, as the spec asks.
structuredContent + outputSchemaEvery tool except echo, color_swatch and read_file
imagecolor_swatch — a valid PNG, base64-encoded
resource_linksearch_web, one per result
embedded resourceread_file
Older revisions get the catalog in their own vocabulary. 2025-03-26 has no outputSchema, structuredContent or resource_link — the JSON text block remains and search_web's text keeps the URLs — and its title moves into annotations.title. 2024-11-05 has neither annotations nor titles.

Resources

resources/list serves four fake files, and resources/templates/list three templates over the same world as the tools.

URITypeWhat it is
file:///etc/app/config.jsonapplication/jsonThe application's configuration file
file:///docs/README.mdtext/markdownThe project's README
file:///data/report.csvtext/csvA small CSV report of scores
file:///images/logo.pngimage/pngA 16×16 PNG logo, read as a base64 blob
TemplateReadsMissing
weather://{city}The current weather, as get_weather reports itweather://Atlantis
user://{user_id}A user profile, as get_user reports ituser://missing
file:///{path}Any file of the fake file system read_file servesfile:///missing
await client.readResource({ uri: "weather://Madrid" });
// { contents: [{ uri: "weather://Madrid", mimeType: "application/json",
//   text: "{\"city\":\"Madrid\",\"temperature\":14,…}" }] }
A missing resource is -32002 up to 2025-11-25 and -32602 from 2026-07-28, as each revision defines it. On 2026-07-28 the list results and resources/read carry cache hints (ttlMs, cacheScope), and resources/read requires the Mcp-Name header. Subscriptions to these resources are covered in Subscriptions.

Prompts

prompts/list and prompts/get serve three prompts. Unknown prompts and missing, unknown or non-string arguments are -32602.

PromptArgumentsAsks for
summarizetext, style?: brief · detailedA summary of a text
code_reviewcode, language?A review of a piece of code
weather_reportcityA short weather report, with the city's weather embedded as a resource. "Atlantis" is never found
await client.getPrompt({ name: "weather_report", arguments: { city: "Madrid" } });
// { description: "A weather report for Madrid", messages: [
//   { role: "user", content: { type: "text", text: "Write a short, friendly weather report for Madrid from this data:" } },
//   { role: "user", content: { type: "resource", resource: { uri: "weather://Madrid", mimeType: "application/json", text: "…" } } } ] }

On 2026-07-28, prompts/get requires the Mcp-Name header. Before 2025-06-18, display titles are left out.