Database MCP server for Claude and ChatGPT Every team gets answers from your data. No SQL. No waiting for reports.
Connect your database once. Your team asks Claude or ChatGPT in plain English and gets answers, charts and dashboards from live data. Read-only, limited to the tables you choose, with every question on record.
Free plan with a sample dataset · Claude connected in about two minutes · Book a demo
-
Plain English
in the AI tool your team already uses
-
Dashboards on request
charts and summaries from live data
-
Your definitions
the agent reads your rules first
-
Safe by design
read-only, every query on record
Works with the AI tools your team already uses, on the databases you already run
- Claude
- ChatGPT
- Cursor
- VS Code
- Claude Code
- + any MCP client
What changes
Your people stop waiting for data and start using it
Most questions about the business already have an answer sitting in your database. Datablare lets the AI tools your team uses reach it, safely.
-
Answers without the report queue
Sales, finance and operations ask in plain English, in the AI tool they already use. The agent reads your live tables and answers, instead of a request joining the report queue.
-
Dashboards from one sentence
Ask for a sales dashboard, a churn view or a stock report, and Claude builds it from your data: the numbers, the charts and what stands out.
-
One version of the numbers
Write down once what “active customer” or “net revenue” means. Every agent reads your rules before it writes SQL, so answers use your definitions, not a guess.
-
A yes from security
Read-only, enforced twice. Only the tables and columns you allow. Every question and query on record. The controls a security review asks for, built in.
Use cases
One request. A dashboard built from your database.
Each of these was built by Claude from a short brief, querying a demo database through Datablare. Every query is read-only and on record.
- E-commerce
E-commerce sales and retention dashboard
₹5.21M revenue from 11,897 orders this year, a 57.9% repeat-customer rate, and which channels bring the customers who spend most.
- SaaS
SaaS MRR, churn and plan mix
$458.7K MRR across 1,760 accounts, 12.8% logo churn over 12 months, MRR by plan and region, and churn month by month.
- Healthcare
Hospital operations and readmissions
10,078 encounters this year, a 30-day readmission rate of 11.1% and an average stay of 5.4 days, broken down by department.
- Banking
Card transaction risk monitor
41,022 card transactions worth ₹16.73M, a 2.46% failure rate, and where failures concentrate by risk level and merchant.
- Manufacturing
Production output, yield and downtime
4.64M good units at 96.9% yield and 84.9% plan attainment, with downtime by line and reason pointing at Line B2.
- Logistics
Carrier delivery performance
24,000 shipments, 80.2% on time and 46 hours from pickup to delivery on average, with on-time rate and cost by carrier and lane.
Before and after
Your team already asks AI about company data. Now it can do it properly.
Today
- Ask the data team, then wait in the queue for a report
- Export a spreadsheet and paste it into ChatGPT
- Each team works out its own version of the numbers
- No record of who looked at which customer data
With Datablare
- Ask in Claude or ChatGPT and get the answer from live data
- Get the chart or the whole dashboard from one request
- Every answer uses the definitions you wrote once
- Every question, and the SQL behind it, on record
Team by team
Chat with your database, whatever your job
The questions people ask on day one. Try them on the e-commerce sample in the Free plan.
-
Sales
Know what sells and who comes back
- “Which products sell the most this quarter?”
- “How many customers ordered more than once?”
-
Finance
Revenue and order value straight from the source
- “What is revenue by category this year?”
- “How did average order value move each month?”
-
Operations
Spot problems while there is time to act
- “Which products are running low on stock?”
- “Which sizes and colours sell best?”
-
Leadership
A read on the business, whenever you need it
- “Build me a one-page dashboard of how we did this month.”
- “How many orders did we get each month?”
How it works
From database to answers in four steps
- Step 1
Step 1
Connect a database
Add PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, ClickHouse or Snowflake, directly or through an SSH tunnel. Or try the e-commerce sample in one click.
- Step 2
Step 2
Choose what agents read, and what it means
Pick the tables each project exposes, hide sensitive columns and give people a role. Add descriptions, rules and example questions so agents use your definitions.
- Step 3
Step 3
Connect your AI tool
Copy the project’s MCP link into Claude, ChatGPT, Cursor, VS Code or Claude Code. Sign in to Datablare and click Allow.
- Step 4
Step 4
Ask, and see every query
Ask in plain English. The agent writes SQL, Datablare checks it and runs it read-only, and Audit records who asked and what ran.
Answers in your terms
Teach it your business once. Every answer uses it.
To an AI, amt_c, is_tst and status = 3 are just
names. Write what your tables and columns mean, and the rules that decide which rows count. Every agent reads it before it writes SQL.
- Descriptions for tables, columns and the links between them
- Rules applied in every query: units, test data, which rows count
- Context for the whole project, such as your fiscal year
- Suggest with AI drafts it for you, from names and types only
orders · 6 columns
amt_c 314218 An amount in dollars? Order total in cents. Divide by 100 for dollars. status 3 Three… what? 1 = paid, 2 = shipped, 3 = refunded. is_tst true Some kind of flag Test order placed by our team. Leave it out of every report. cust_typ 'B' B for… best customer? B = business buyer with a tax ID, C = consumer. zone_cd 'NE' Nebraska? North-east? NE = our North-East sales region. created_at 2026-10-04 18:40 Local time, probably Stored in UTC. Count days in Eastern Time. Written once, read by every agent. Datablare sends these descriptions with the schema before the agent writes SQL, in Claude, ChatGPT, Cursor or VS Code.
Illustrative example. Your columns, your descriptions.
For your security team
Checked three times. Recorded once.
Every question passes Datablare’s checks on the SQL, the tables and the columns before anything touches your database, and all of it is written down.
Illustration. A person asks in Claude, ChatGPT or Cursor: who are our top customers this month? The question reaches Datablare, which checks it three times: the SQL guard confirms it is a single read-only SELECT, the tables are on the project's allowed list, and the hidden phone column is not requested. The query runs on the database, which also keeps it read-only. Three rows come back without the phone column, and the call is added to the audit log. Then someone asks to delete last month's test orders: the DELETE statement is refused at the SQL guard, never reaches the database, and the refusal is recorded in the audit log too.
Your AI tool
Claude · ChatGPT · Cursor · VS Code
- “Who are our top customers this month?” Answered
- “Delete last month’s test orders.” Refused
Datablare checks
- SQL guard: read-onlyOne SELECT. Writes are refused.
- Allowed tables onlyorders, customers are on the list
- Hidden columns stay outphone is hidden by your admin
DELETE stopped at the SQL guard. It never reaches the database.
Your database
Runs the query, and the database itself keeps it read-only too: a second lock behind the guard.
Read-only, twice
Rows back to the agent
| Customer | Orders | Spent | Phone |
|---|---|---|---|
| Emma Clarke | 14 | $12,480 | ••• |
| Lucas Martin | 11 | $9,845 | ••• |
| Aisha Khan | 9 | $8,620 | ••• |
Phone is hidden, so it is never returned. Results pass through and are not stored.
Audit log
- 10:42Delete test orders Refused
- 10:42Top customers this month Got data
- 10:37Revenue by category Got data
Who asked, the SQL that ran, the tables it touched, how long it took.
The controls
Read-only database access for AI, with the controls security asks for
-
Read-only, enforced twice
Datablare’s SQL guard lets only a single read through, and the database itself keeps the query read-only. Agents cannot insert, update, delete or drop.
-
Down to the column
Choose which tables each project exposes and hide individual columns: salary, national ID, phone, diagnosis. Hidden means never returned.
-
Every query on record
Audit shows who asked, the question, the SQL that ran, which tables it touched and how long it took. Export it as CSV on paid plans.
-
Works with the AI tools you use
One MCP link per project for Claude, ChatGPT, Cursor, VS Code, Claude Code or any MCP client. People sign in with their own account.
Access control
Same question, different access
Every person reaches the data through their own sign-in, so the agent can only read what that person is allowed to. Switch roles to see what comes back.
Same question in Claude: “Who were our top customers last quarter, and how do I reach them?”
Owner or manager: the agent can read every source in the project. Hidden columns stay hidden for everyone, owners included.
Analyst given the Sales source only: rows that live in the ERP source are never returned to their agent.
Viewer: their agent cannot read data in this project. In Audit they see that calls happened, not the questions, SQL or answers.
Illustration of real behaviour. Names and figures are examples.
Multiple databases
One MCP link. All your databases.
Connect PostgreSQL, MySQL, Snowflake and more to one project. Your team adds a single link to Claude or ChatGPT and asks questions that span them, with one access policy and one audit log.
Illustration. Claude, ChatGPT and Cursor all connect to one Datablare MCP link for a project. The project has three databases: PostgreSQL holding orders, MySQL holding the CRM, and Snowflake holding finance data. When someone asks a question that spans them, the agent sends one read-only query to each database it needs. Datablare checks every query, runs it on the right database, and records it in the audit log. The agent then combines the results into one answer.
“Which customers in the CRM spent over $10,000 last month?”
The agent queried CRM and Orders separately and combined the results. Both queries are in Audit, under the person who asked.
Illustration. One query runs on one database; the agent combines the answers. Names and figures are examples.
Build, borrow or use Datablare
What a governed MCP gateway saves you building
You can wire an agent to a database with an open-source MCP server in an afternoon. Making it safe for a whole company is the part that takes months.
| What you need | Datablare | Open-source MCP server | Build it yourself |
|---|---|---|---|
| Read-only enforcement | Yes: SQL guard, and read-only at the database too | Partly: Typically depends on the login you give it | Partly: You build and test it |
| Choose tables and hide columns | Yes: Per project, in the app | Partly: Typically via database grants you maintain | Partly: You build it |
| Per-person roles and sign-in | Yes: Manager, analyst, viewer; OAuth sign-in | No: Typically one shared connection string | Partly: You build it |
| Audit log of every query | Yes: Who asked, the SQL, tables, timing; CSV export | No: Typically local logs only | Partly: You build and store it |
| Works with Claude, ChatGPT, Cursor, VS Code | Yes: One remote MCP link per project | Partly: Often a local install on each laptop | Partly: You build the MCP server |
| Context for the agent (descriptions, rules) | Yes: Modeling: tables, columns, rules, example questions | No: Typically raw schema only | Partly: You build it |
| Hosting and billing | Yes: Run for you, priced in US dollars; your own cloud for Enterprise | Partly: You host and run it | Partly: You host, run and maintain it |
Generic comparison of common approaches, not of any named product. Open-source servers vary; check the one you are considering.
Deployment
Our cloud, or yours
Datablare Cloud
Free, Team and Business
- Run by us, priced in US dollars (or rupees in India)
- Connect databases directly or through an SSH tunnel
- Credentials encrypted; query results never stored
Your own cloud
Enterprise
- For regulated data or data-residency requirements
- Negotiated limits and a deployment scoped with you
- Help with your security review
Built for regulated data
Helps you meet the privacy laws your customers ask about
Datablare gives you the controls auditors look for when AI tools touch personal data. Compliance stays yours; these make it much easier to show.
Helps you meet
- GDPR
- HIPAA
- CCPA/CPRA
- PDPL
-
Read-only
Enforced by the SQL guard and by the database itself.
-
Column-level control
Hide personal data columns from every agent.
-
Full audit trail
Who asked, the SQL, the tables touched, when.
-
Results never stored
Rows pass to your AI tool and are not kept. Enterprise can run in its own cloud.
FAQ
Questions people ask first
Something else? Book a call with Kamal or write to [email protected].
What is a database MCP server?
MCP (Model Context Protocol) is the open standard AI tools like Claude, ChatGPT, Cursor and VS Code use to call outside tools. A database MCP server gives those tools a way to read a database: list tables, write SQL and get rows back. Datablare is a hosted, governed one: it checks every query, keeps it read-only, limits it to the tables and columns you choose, and records it.
Can the AI change or delete data?
No. Every query is checked by Datablare’s SQL guard, which only lets a single read through. It then runs with the database itself keeping it read-only: a read-only session on PostgreSQL, MySQL, MariaDB, Oracle and ClickHouse, and a transaction that is always rolled back on SQL Server and Snowflake. Inserts, updates, deletes and schema changes are refused.
Does the AI see my data?
It sees the rows it asks for, and only from the tables and columns you allow. Hide a column (a phone number, salary or national ID) and it is never returned to any agent. Results pass through Datablare to the AI tool and are never stored by Datablare; the AI provider handles them under its own terms.
How does the agent know what our tables mean?
From the context you write in Modeling: a friendly name and description for each table, rules the agent must follow (units, exclusions, which rows count), example questions, column descriptions, what links between tables mean, and context for the whole project such as your financial year. Datablare sends it to the agent with the schema, before it writes SQL, in every AI tool. Context guides how the query is written; it never widens what the agent may read.
Can we use AI tools on customer data under privacy laws like GDPR?
Privacy laws such as GDPR and CCPA/CPRA do not ban AI tools. They expect you to use only the personal data you need, protect it, work with providers under proper terms, and be able to show who accessed what. Datablare helps with the parts that involve AI agents: hide personal data columns, read-only access with per-person sign-in, and an audit record of every query. Compliance stays with your organization; check your use with your lawyer.
Which databases and AI tools work with Datablare?
Databases: PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, ClickHouse and Snowflake, directly or through an SSH tunnel. AI tools: Claude, ChatGPT, Cursor, VS Code (GitHub Copilot), Claude Code and any other MCP client that supports remote servers.
How do people connect their AI tool?
Each project has one MCP link. Paste it into the AI tool, sign in to Datablare and click Allow: that OAuth sign-in is the default for every client. A personal key is the fallback for tools that cannot sign in (ChatGPT is sign-in only). Analysts and above can approve a connection themselves; viewers send a request to an admin.
Can one MCP link query multiple databases?
Yes. Every database you add to a project is reachable through that project’s single MCP link, and they can be different engines: say PostgreSQL for orders, MySQL for the CRM and Snowflake for finance. The agent lists the project’s data sources and sends each read-only query to the database it belongs to; for a question that spans two, it queries each and combines the results. One SQL statement runs on one database, so there are no cross-database joins. You can still limit a person to particular databases on the same link.
Where is Datablare hosted, and what does it store?
Datablare Cloud is hosted in India. It stores your connection details (encrypted), the context you write about your tables, and an audit record of each call: who asked, the question, the SQL, the tables touched and timings. It never stores query results.
Can I try it without connecting my own database?
Yes. The Free plan includes a one-click e-commerce sample dataset, already described, so you can connect Claude and ask real questions in a couple of minutes before you connect anything of your own. The sample does not count towards your data sources.
Can we run Datablare in our own cloud?
For Enterprise customers, deploying in your own cloud is a conversation we scope together, usually for regulated data or in-country hosting requirements. Book a call or email us.
Incubated at IIT Mandi Catalyst
Supported by Startup Himachal
Give every team answers from your data.
Start free with the e-commerce sample or your own database. Connect Claude in about two minutes and ask your first question.
30 minutes with the founder. Or WhatsApp / [email protected]

