# Mostly Right Mostly Right publishes open, machine-readable datasets as immutable Apache Parquet versions: time series, benchmarks and reference data for analysis and machine learning. Every dataset carries a documented column schema, column profiles, an explicit licence and a named publisher. There are three tiers. Reading metadata, schemas and sample rows needs no account, and neither does a bounded query of up to 100 rows a page (POST /api/public/tables/{table_id}/query). Querying a whole table — up to 10000 rows a page, paged to the end — and downloading a full snapshot need a free account and an API key from Settings → Access; the first request connects the dataset to your workspace. Building a new dataset — writing a recipe, running it and publishing the result — needs a paid subscription on that workspace. - Dataset directory: /datasets - Every public dataset and table, in one file: /llms-full.txt - One dataset as Markdown: /datasets/{slug}/llms.txt - Public Dataset API: https://mostlyright.md/docs/reference/public-dataset-api/ - Use a public dataset: https://mostlyright.md/docs/guides/use-public-datasets/ - Quickstart for agents: https://mostlyright.md/docs/guides/for-agents/ - Connect an AI tool with MCP: https://mostlyright.md/docs/guides/mcp/ - MCP server: /api/mcp (server card: /.well-known/mcp.json) - Agent skills: /.well-known/agent-skills/index.json - How to build a dataset: /.well-known/agent-skills/mostlyright-build/SKILL.md - OpenAPI: /api/v2/public/openapi.json Use the traditional HTTP API for application code. GET /api/v2/public/datasets searches the catalogue and needs no key. GET /api/v2/public/tables/{tableId}/schema and GET /api/v2/public/tables/{tableId}/preview need no key. GET /api/v2/public/tables/{tableId}/current returns application/vnd.apache.parquet. POST /api/v2/public/tables/{tableId}/query returns bounded JSON. The last two take a workspace key as x-api-key: mr_use_…. Do not assume arbitrary SQL is available. Agents can use the MCP server at /api/mcp instead: Streamable HTTP, protocol 2025-11-25 (2025-06-18, 2025-03-26 and 2024-11-05 also negotiate), stateless, no session id. Server card: /.well-known/mcp.json. search, fetch, search_datasets, get_dataset, list_tables, get_table_schema, sample_rows, get_download_instructions and get_access_instructions answer with NO credential; search and fetch are the two-tool connector contract, where search returns {id, title, url} per dataset and fetch returns that dataset as Markdown. query_table, catalog_search and list_connected_datasets take a read-only mr_use_ workspace key as Authorization: Bearer, or an OAuth connection. connect_dataset changes the workspace and needs the OAuth connection. catalog_search is a different question from the rest: it searches a sealed snapshot of public data SOURCES — one provider, Data.gov, partially swept — rather than the datasets published here. The same server also builds datasets, on a paid workspace, over an OAuth connection carrying the datasets:build scope: create_dataset, update_dataset, list_my_datasets, get_my_dataset, register_recipe, propose_revision, replay_run, start_run, confirm_run, approve_full_run, cancel_run, get_run, run_events, list_runs, query_run, run_artifacts, get_artifact_download, write_note, promote_table, get_table, diagnose_table, list_source_credentials, normalize_reader_options, start_source_inspection and get_source_inspection. An mr_use_ key cannot build. Source credentials are pasted by a person at /dashboard/settings/secrets and referenced from a recipe by name; no secret passes through MCP. The loop is written out at /.well-known/agent-skills/mostlyright-build/SKILL.md and served as the mostlyright://build-guide resource. Results should cite the dataset, table, immutable version, content digest, and query time. ## Build a dataset Building runs from the command line as well as over MCP. Install it with pip install mostlyright-data, which puts the mr-data command on the path, then run mr-data login once per machine. The first mr-data command of any kind also places the mr-data-build agent skill at ~/.claude/skills/mr-data-build, or under ~/.codex/skills for Codex, and leaves an edited copy alone; MOSTLYRIGHT_SKILL_AUTOINSTALL=0 switches that off. That skill drives the CLI. The MCP build skill is a separate document: /.well-known/agent-skills/index.json lists it, and it is served at /.well-known/agent-skills/mostlyright-build/SKILL.md and as the mostlyright://build-guide resource. ## Documentation Start here. - https://mostlyright.md/docs/ — what Mostly Right builds, and which page to read next. - https://mostlyright.md/docs/start/install/ — install mr-data and sign in. - https://mostlyright.md/docs/start/first-dataset/ — build one dataset end to end. - https://mostlyright.md/docs/start/how-it-works/ — the lifecycle: dataset, recipe, run, sealed table version, refresh. - https://mostlyright.md/docs/guides/for-agents/ — the curl ladder from a catalogue search to a cited result. Build a dataset. - https://mostlyright.md/docs/build/probe-sources/ — ask one source one question before writing a recipe. - https://mostlyright.md/docs/build/write-a-recipe/ — the recipe document, with a complete example. - https://mostlyright.md/docs/build/run-and-inspect/ — run under ceilings, follow the events, read the rows. - https://mostlyright.md/docs/build/publish-and-refresh/ — what a succeeded run publishes, and how refresh cadence is recorded. - https://mostlyright.md/docs/build/credentials/ — store a source credential by name and reference it from a recipe. - https://mostlyright.md/docs/build/live-streams/ — record a public wss venue and build a table from the recording. - https://mostlyright.md/docs/build/collections/ — one source that fans out over many pages. - https://mostlyright.md/docs/build/documents/ — documents and files as rows. - https://mostlyright.md/docs/build/troubleshooting/ — read a failed run and fix what it names. Recipe examples, each one a complete document. - https://mostlyright.md/docs/recipes/ — the index. - https://mostlyright.md/docs/recipes/city-temperatures/ — daily city temperatures. - https://mostlyright.md/docs/recipes/csv-snapshot/ — one CSV file to one table. - https://mostlyright.md/docs/recipes/json-api/ — a JSON API with a records pointer and pagination. - https://mostlyright.md/docs/recipes/html-collection/ — many HTML pages to one table. - https://mostlyright.md/docs/recipes/document-extraction/ — a PDF or other document projected into rows. - https://mostlyright.md/docs/recipes/weather-grib/ — weather GRIB2 through the scientific reader. - https://mostlyright.md/docs/recipes/websocket-stream/ — a WebSocket venue to a stream table. - https://mostlyright.md/docs/recipes/multi-source-join/ — two sources joined, with checks. - https://mostlyright.md/docs/recipes/market-settlement-join/ — settlement rules from two prediction venues joined on one event. - https://mostlyright.md/docs/recipes/forecast-vs-observation/ — a weather forecast joined to the observation it predicted. - https://mostlyright.md/docs/recipes/snapshot-window/ — a listing replaced as one snapshot on each refresh. - https://mostlyright.md/docs/recipes/many-station-weather/ — twenty station feeds unioned into one hourly table. - https://mostlyright.md/docs/recipes/stream-to-bars/ — recorded trade frames aggregated into one-minute bars. Reference. - https://mostlyright.md/docs/reference/cli/ — every mr-data command, one line each. - https://mostlyright.md/docs/reference/cli/auth/ — login, auth, whoami, open, clarify. - https://mostlyright.md/docs/reference/cli/dataset/ — dataset and cover. - https://mostlyright.md/docs/reference/cli/recipe/ — recipe, probe, catalog, connections, keys. - https://mostlyright.md/docs/reference/cli/run/ — run, runs, status, watch, checks, receipt, note, diff, verify. - https://mostlyright.md/docs/reference/cli/read/ — peek, query, download, parts. - https://mostlyright.md/docs/reference/cli/table/ — table, promote, reschedule, pin, unpin, demote. - https://mostlyright.md/docs/reference/cli/stream/ — stream venues and recordings. - https://mostlyright.md/docs/reference/recipe/ — the recipe document, field by field. - https://mostlyright.md/docs/reference/sources/ — source kinds and connectors. - https://mostlyright.md/docs/reference/readers/ — readers. - https://mostlyright.md/docs/reference/transforms/ — transforms. - https://mostlyright.md/docs/reference/checks/ — checks. - https://mostlyright.md/docs/reference/units/ — units and vocabulary. - https://mostlyright.md/docs/reference/run-states-and-errors/ — run states, events, refusal and worker codes. - https://mostlyright.md/docs/reference/limits/ — limits and ceilings. Reading a dataset somebody else published is covered by the three pages named at the top of this file: the Public Dataset API, Use a public dataset, and Connect an AI tool with MCP.