A database for AI assistants, over MCP

A database your AI can actually use.

Connect Claude, Cursor or any MCP assistant once. It loads your files and tables, keeps its own working data, queries millions of rows exactly, and sends the reports you used to build by hand.

Request an invite Connect in two minutes

Works with Claude Desktop, Claude Code, Cursor, VS Code, Python and Jupyter.

Your AI assistant, connected to the public demo

Using mcpdata, show total revenue per month from Demo.Orders, oldest month first.

list_tables() → 3 tables: Demo.Orders, Demo.Products, Demo.Customers

describe_table(table_name="Demo.Orders") → 15 columns, 10,005,000 rows

query(table_name="Demo.Orders",...) → 36 rows

Revenue by month (the demo's synthetic orders)

36 monthly totals from one query over every order, run on 2026-10-29. See the recipe.

Three steps. Then just ask.

Connect

One line in your assistant's MCP settings, with your key.

claude mcp add --transport http mcpdata https://[you].mcpdata.org/mcp --header "Authorization: Bearer $KEY"

Load

Your AI creates tables and appends rows itself. Or point cdb-localdata at a folder of CSV, Parquet, ORC or Excel files and they become tables, with a dated snapshot each day if you want one.

Report

Exact answers, not guesses. Save a query as a view, turn it into a chart or pivot, and have it emailed as HTML, CSV or Excel on any schedule.

Move data in. Keep it. Hand it on.

Files too big for a chat window live here once, and your AI queries them as often as it likes. Only the answers enter the conversation.

Files from your computer

CSV, Parquet, ORC and Excel folders become tables. Your files stay on your machine until you sync.

Loads of any size

Big files are sent in parts automatically. If a part fails it's simply retried, with no duplicate rows.

Your AI's own tables

Agents keep logs, snapshots and state in their own agent.* tables, between conversations.

Every version kept

Each load is kept as a version. Ask what a table looked like on a past date, or what changed.

Reports that run themselves.

Exact numbers. Your AI queries every row, not a sample, so totals match every time.
Saved views with parameters. One approved query, run the same way each time.
Charts and pivots. Bar, line, area, scatter, pie, histogram; report files open offline.
On a schedule. Emailed as HTML, CSV or Excel, with a history of every run.
Time travel. Ask what a table looked like on a past date with AS OF.

How do I pivot data by month and region with AI? (the recipe)

order_month	APAC	Americas	EMEA
2025-01	31,308,523	65,707,869	59,462,781
2025-02	28,851,954	60,315,307	54,646,255
2025-03	32,284,182	67,671,765	61,718,086
2025-04	31,132,655	65,476,850	59,046,685
2025-05	33,213,488	69,285,388	62,213,675
2025-06	33,667,736	69,698,346	63,292,061
2025-07	36,987,809	76,946,704	69,931,726
2025-08	35,804,577	74,864,596	67,624,060

From the public demo's synthetic orders, verified 2026-10-29.

Questions people ask their AI

Every answer is a recipe we run against a live instance, with the date it last passed.

Browse all 50 answers

Your data, your key.

Encrypted with its own key

Each instance's data is encrypted at rest with a key of its own. Delete the instance and the key goes, so every copy becomes unreadable.

Sign-in with 2FA

Read, ReadWrite and Admin roles. API keys that expire and work only from the addresses you choose.

Everything audited

Every sign-in, key and change is logged. Revoke an AI's access in one click.

Shared or dedicated

Start on a shared instance; move to your own server with the same API.

Give your AI somewhere to put the data.

Free shared instances are opening by invitation. Paid plans: coming soon.

Request an invite