Catalogian MCP¶

A spreadsheet answer is only as current as the evening somebody re-exported it. The useful question is not what the file says, it is what changed in it since the last time anyone looked.

Server type: Remote (Streamable HTTP)
Endpoint:    https://api.catalogian.com/v1/mcp
Auth:        anonymous session token, record key or account key
Tools:       18 measured, all read-only
Record:      a CSV, TSV or JSONL source, versioned on every ingest
Delta:       per row, new, changed or deleted, with the key
Pricing:     free 2 records, $4 and $12 per record monthly

What it does¶

Catalogian turns a data source into a record: a CSV, TSV or JSONL file, fetched from a URL or uploaded, keyed on a column, indexed, and versioned. Every ingest becomes an edition, and every edition produces a delta event that names the rows added, changed and deleted. Each record carries its own MCP endpoint, so an agent gets the record and the tools in one address.

The eighteen tools split into three groups.

Group	Tools
Shape and sample	`snapshot_schema`, `profile_snapshot`, `sample_snapshot`
Read and query	`get_snapshot_rows`, `query_snapshot`, `filter_snapshot_rows`, `search_snapshot`, `download_snapshot`, `download_filtered_snapshot`
History and change	`list_editions`, `get_edition_diff`, `get_delta`, `get_delta_rows`, `compare_snapshots`

list_records, get_record_by_slug, get_health and list_tools complete the set. The aggregations run server side rather than in the client, so query_snapshot answers a group_by over a large record without the model reading every row.

The onboarding path is the part worth noticing. No account is needed to start: POST /v1/ephemeral/ingest with a URL returns a session token, and that token is the credential for all eighteen tools for twenty four hours, three ingests and fifty tool calls, with 10,000 rows indexed per record. A record that matters is then claimed and becomes a watched record with a schedule and a delta feed.

What this pass measured¶

A driven session against the live endpoint, not a documentation read.

An anonymous ingest of a public business dataset (https://data.cityofchicago.org/resource/ydr8-5enu.csv?$limit=300&$order=id) returned a session and a source in one call: rowCount 300, indexedRows 300, rowsSeen 300, keyField id, format csv, tier free, truncated false, and a first delta of 300 new rows.

initialize answered HTTP 200 with serverInfo catalogian 1.1.0, and tools/list returned all 18 tools with schemas in 28,237 bytes. list_records confirmed the session is scoped to one record and returned the meter: 49 of 50 tool calls remaining, session expiry the next day.

snapshot_schema describes the record rather than dumping it: total rows 300, sampled 100, and 117 fields each with a null rate, a type hint and example values. get_snapshot_rows returned rows as rowKey plus data, which is the shape a model can cite.

Four calls on the same record, all answered:

query_snapshot with group_by on permit_type returned the whole distribution in one object: 299 rows in the express permit program and 1 in signs.
filter_snapshot_rows with contains SIGNS returned 1 matching row and stated that all matches were shown.
search_snapshot for electric returned rows ranked by relevance rather than by position.
download_snapshot failed. The presigned export answered The request signature we calculated does not match the signature you provided, an S3 signing error on the vendor side, so the export path is broken while the query path works.

The change machinery was then driven with a second ingest of the same record at a different slice ($offset=100). It landed on the same record, ingestCount moved to 2, and the delta was exact: 100 new, 100 deleted, 200 unchanged, 300 total. list_editions showed edition 1 with kind: first and edition 2 with kind: revision carrying its own deltaEventId. get_edition_diff from edition 1 to edition 2 returned row level results with changeType: new and the full field set for each row, which is the before and after an operator needs.

One inconsistency is recorded rather than smoothed. compare_snapshots on the same two editions answered Added 100, Removed 0, Modified 0, while get_delta on the same edge answered deletedCount: 100. The two tools disagree about deletions, so a deletion test should be run through get_delta or get_edition_diff until the vendor reconciles them.

Where it sits in this catalogue¶

Nothing else here treats a file as a versioned object. Settra MCP is the nearest neighbour and it goes the other way: it syncs Google Sheets, CSV, Excel and Parquet into PostgreSQL behind a semantic layer, so an agent queries governed definitions that a data team approved. That is the right shape when several people share one warehouse. Catalogian is the shape when one operator has one file that keeps arriving, wants to know what changed in it, and does not want to stand up a database to find out.

The distinction matters in practice. Settra answers "what does the approved definition of revenue say". Catalogian answers "which rows in this week's export are not in last week's, and what was the previous value of the field that moved". Those are different questions, and only one of them is about a single file.

What this pass did not establish¶

The key field is a closed list, not any unique column. The ingest refused a clean state population CSV with NO_KEY_FIELD, naming the accepted keys: id, product_id, item_id, variant_id, sku, item_number, upc, ean, gtin, isbn, asin, mpn. A dataset keyed on a name, an email or a region is not addressable without renaming a column first.

A second defect was met on a vendor sample file. The WooCommerce sample products CSV opens with a byte order mark on its header, and Catalogian took the key field as the BOM prefixed ID, then refused the revision with KEY_FIELD_MISSING because the established key could not be found in headers that plainly contain it. Files exported from Excel or from a vendor sample can carry that mark, so a failed revision on a file whose columns look correct should be read as a BOM problem, not as a data problem.

Scheduled checks, webhooks and the claimed hourly cadence were not driven, because they need a claimed record and a paid rung. The free anonymous path has no schedule at all: it is a snapshot and a manual revision. The row cap was not reached, so sampling behaviour above 10,000 rows is the vendor's statement. download_filtered_snapshot, sample_snapshot, profile_snapshot and get_delta_rows were listed and not called.

FAQ¶

Is it free to try?¶

Yes, and without an account. One call creates an anonymous record and returns a token that opens all eighteen tools for twenty four hours, three ingests and fifty tool calls, with 10,000 rows indexed.

What counts as a record?¶

One source file, CSV, TSV or JSONL, capped at 520 MB per ingest with gzip read directly, keyed on one column, versioned on every ingest. Each record gets its own MCP endpoint.

Do I need to write SQL to ask a question?¶

No. query_snapshot runs distinct, count, group_by, min, max, avg, sum and filter server side, and filter_snapshot_rows joins conditions with AND. The measured group_by returned one object with every category and its count.

How do I know what changed since last time?¶

get_delta returns the counts and the keys per event, get_delta_rows returns the rows, and get_edition_diff returns field level before and after between two editions. On the measured revision the delta was 100 new and 100 deleted against 200 unchanged.

Does it work in Claude, ChatGPT and Cursor?¶

Yes. The record page issues a self-contained MCP URL that carries the token, for clients that cannot set an Authorization header, and the docs prefer the header where the client supports it.

What happens when the data does not have one of the accepted key columns?¶

The ingest fails with NO_KEY_FIELD and lists the accepted keys. Rename one column to id or sku and re-ingest.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects