Georgia Civic Data MCP Server

Enumerate every approved Georgia dataset (topic) and the shared dimensions. Each topic entry is a LEAN summary — name/keys, year coverage (year_min/year_max + year_gaps), detail levels + default detail, a has_demographic flag (false = no demographic axis, so there is no all-students demographic row to filter), tags, contract version, and a one-line description — enough to pick a topic; call describe_dataset for its full schema (columns, filters, grain, source, example queries). Each dimension entry carries its primary key, attribute columns, and (for districts) cross-dataset link keys. Call this first to learn what exists — but for a NAMED task (you already know roughly the topic), prefer search_datasets, which returns far fewer bytes than this full catalog.

Keyword search over the catalog metadata (topic names, descriptions, tags, AND column names/descriptions) — the discovery entry point when you don't know the exact topic name. Returns lean topic summaries per hit with a relevance score and which fields matched, plus a `dimension_matches` list when the query also hits a dimension (e.g. 'district'). Most acronyms work; the short ones `ap`/`el`/`ib` are recognized. Follow up with describe_dataset. `limit` caps results (default 20, max 100).

Full schema for one topic: every column (name/type/role/unit/value range/null-meaning), the exact `filters` list with enum values (read this before query_dataset — it is the authoritative set of filter keys), the FK→dimension join shape (`foreign_keys`), example queries, usage, limitations, null semantics, tags, and `schema_hash` (for cache/drift detection). The top-level `key_metric` names the single headline column most answers want; each column carries `key_metric_grain_contributor` (a grain axis the key metric is only comparable within — pin or group by it) and `metric_component` (numerator/denominator of a rate/average metric). `recommended_query` gives the safe default query shape (key metric + filters to pin + required single-selects) plus a `ranking` recipe for top/bottom-N asks; `filter_hints` lists paired filters; each categorical filter carries `has_total` / `requires_single_value`. Pass `verbosity='schema'` for a much smaller payload that drops the prose (description/usage/limitations/example queries/column descriptions) but keeps every field needed to compose a correct query — use it when you only need the filter keys and enums; prefer the default 'full' before reporting conclusions (the limitations prose carries the caveats). On an unknown topic returns a self-describing error listing available topics + a 'did you mean' hint. `main_topic` defaults to 'education'; pass 'census' for Census topics or 'immigration' for immigration topics.

Schema for one dimension (districts / schools / counties / demographics): the (possibly composite) primary key, the attribute columns a join attaches, the cross-dataset `link_keys` (e.g. districts.district_census_id → Census via the crosswalk — a 5-digit school-district code, NOT a county FIPS), and demographics `semantics` (within a category the values are mutually exclusive; `all` is the denominator). Read this before writing a link_query join.

Run a cross-dataset / cross-topic analytical SQL query that the per-topic query_dataset filters can't express — e.g. join education, Census, or immigration facts to a dimension (or another dataset) on shared geography (immigration and Census county topics share county_fips directly). READ-ONLY, SANDBOXED DuckDB: one SELECT (or WITH … SELECT); no DDL/DML/COPY/ATTACH/INSTALL/PRAGMA/SET/CALL; you may only read_parquet() the curated gold paths returned by link_tables (call it first and paste the snippets) — querying a file path directly is rejected. Joins use the keys from describe_dimension's link_keys (districts.district_census_id bridges to Census via the crosswalk — it is a school-district code, not a county FIPS, so a district is not 1:1 with a county). Results are row- and byte-capped and time-limited; `truncated` flags when capped — add aggregation or a tighter WHERE rather than dumping rows. NULL means suppressed, not zero. On a violation you get a self-describing error naming the offending token/path.

List the tables link_query can read (curated gold paths only) and the join keys that bridge facts → dimensions → Census geography. Call this BEFORE writing a link_query. Two-tier to stay context-cheap: with NO arguments it returns a LEAN index — every table's name, grain, detail levels, default `read_parquet(...)` snippet, and join keys (enough to pick tables and write a single-detail join). To get every column and a snippet per detail level for the few tables you actually need, call again with `tables=["<name>", ...]` (a `name` from the index, e.g. 'education/gosa/attendance' or 'attendance', or a dimension like 'districts'). Paste the `read_parquet(...)` snippets verbatim into your SQL — they are exactly what the sandbox accepts.

Compute a grouped aggregate over one topic — the aggregation-first path. `agg` is one of avg/sum/min/max/count/weighted_rate; `metric` is a metric column (DEFAULTS to the topic key_metric; ignored for count); `group_by` is a list of grain columns (year, FK codes like district_code or county_fips, or categoricals — see describe_dataset). `weighted_rate` computes a true population-weighted SUM(numerator)/SUM(denominator) for a rate key metric (when the contract declares the components) — prefer it over `avg` for a rate across multiple places/years, since `avg` means the per-row rates and ignores population. Supports the same `filters` / `year` / `year_min`-`year_max` / `detail` as query_dataset, plus `order_by`+`order` for top-N (order_by 'value' for the aggregated column; NULL cells sort LAST in either direction). Returns one small row per group with `<metric>_<agg>` (or `row_count`) plus coverage diagnostics (input_rows / non-null counts) so suppression is visible; `aggregation_scope` flags whether rows are source-published at this grain or recomputed from a finer detail (prefer source-published — see the advisory). Aggregates SKIP NULLs and NULL means SUPPRESSED not zero. No raw SQL: all identifiers are contract-allowlisted.

Return the authoritative ODCS v3.2 data contract for a topic (kind='topic') or a dimension (kind='dimension') so you can consume the machine-readable schema without cloning the repo. fmt='yaml' (default) returns the document verbatim as text; fmt='json' returns it parsed. Only approved topics and loaded dimensions expose a contract.

Query one topic's gold facts with dimension labels joined in (the district/school/county/demographic names come back on every row). `filters` is a dict of column → value or list-of-values: FK codes (district_code, school_code, county_fips, demographic) and any categorical column — see describe_dataset's `filters` for the exact keys and enum values. Use `year` (exact) OR `year_min`/`year_max` (range), never both. `detail` picks the grain (default is the finest available). Returns `rows` plus a `columns` descriptor array (type/role/unit/null-meaning, and `is_key_metric` flagging the headline column) so you interpret values and NULLs correctly — NULL usually means SUPPRESSED, not zero (see null_semantics). The top-level `key_metric` echoes which column is the answer. Use `columns` to project a subset, `include_labels=false` to skip the joined name columns (codes only), and `order_by`+`order` for server-side top-N instead of over-fetching. Pages are small (default 100, max 500); when `truncated` is true a `bulk_export` block points at the REST CSV/Parquet endpoint and the source path for the full pull — do not loop pagination to dump a table. A bad filter returns a self-describing error listing the valid keys/values.

Paginated read of a dimension table — the label lookups (district names, school names, county names, demographic labels). Rows are ordered by the dimension's primary key so paging is stable. Use describe_dimension for the schema and link keys. Small page defaults; `truncated` + a `bulk_export` pointer signal when to pull the full table elsewhere.

List the distinct values of ONE filterable column of a topic — the fast way to learn valid filter values before query_dataset, especially for FREE categoricals and FK codes (district_code/school_code/county_fips/demographic) that carry no enum in describe_dataset (a wrong guess otherwise returns an empty page with no error). `column` must be a filterable column (see describe_dataset's `filters`). Optional `prefix` does a case-insensitive starts-with filter; `limit` caps results (default 50). Enum-bearing columns return their contract enum directly; others run a capped SELECT DISTINCT over the gold data. `truncated` flags when the list is capped.

Resolve a place or demographic NAME or CODE to its stable keys + labels — the right way to turn 'Atlanta Public Schools' / 'Fulton' / a raw code into the district_code / school_code / county_fips / demographic to filter by (a wrong code guess otherwise returns an empty query_dataset page). `kind` is district / school / county / demographic ('Fulton' as kind='county' → the county; as kind='district' → the school district — they are different things). Fuzzy-matches and ranks candidates, flags `ambiguous` when several tie, and reads only the small dimension table (no fact scan).