Your AI gets a front desk, not the keys: how Hutif keeps your data apart from AI

Your AI gets a front desk, not the keys: how Hutif keeps your data apart from AI

Here is how most people connect an AI to their business data today: they copy a database password, an API key or a whole spreadsheet, and paste it into a chat. It works. It also hands a piece of software that guesses for a living the master key to everything that password can open.

Hutif was built on the opposite idea. Your AI should be able to ask questions about your data without ever holding your data's keys. It gets a front desk: a short list of clear things it may ask for, a person at the counter who checks every request, and a door that only you can unlock. This article walks through exactly what sits at that desk, and what never leaves the back office.

1. Your AI never sees your passwords

When you connect a database or an API to Hutif, the credentials go into Hutif once and stay there.

  • Encrypted at rest. Database passwords and API credentials (keys, tokens, Basic auth) are encrypted before they are stored, and the keys that unlock them are kept separately, so stored data alone can't be read.
  • Never shown again. After you save a password, no screen in Hutif will display it back, not even to you. To change it, you type a new one.
  • Never given to the AI. Credentials are decrypted in memory only at the moment Hutif opens the connection on your behalf. None of the tools your AI can call returns them, and they never appear in what the AI is told about the server.

That one decision removes the biggest risk of the paste-it-in-chat approach. A password that never reaches the conversation can't end up in chat history, in logs on someone else's platform, or in a screenshot shared in a team channel.

2. Your AI sees a few tools, not your database

An AI connected through Hutif doesn't get a database connection. It gets a menu. For a database or a spreadsheet, that menu is:

  • list the tables (or sheets),
  • look at a table's columns,
  • read rows, with filters, up to 200 rows per request.

Hutif builds every query itself. Values from the AI are always passed as parameters, never glued into the SQL text, and table and column names are checked against your real schema before anything runs. Queries time out after 60 seconds, so a single question can't hold your database hostage.

Free-form SQL is a separate switch that is off by default. If you ever turn it on, Hutif tells you plainly what it means: any statement, with the full rights of the database user you connected.

APIs get the same treatment. If you upload an OpenAPI description, the AI only sees the operations you picked. Every call is locked to the API's own address, so the AI can't use your server to reach other websites, and private network or cloud-metadata addresses are blocked.

3. Read-only until you say otherwise

Every new Hutif server is read-only. When writing is off, the tools that change data are not just disabled, they are not even shown to the AI. It can't call what it doesn't know exists. For an API, read-only means only GET requests.

When you do want your AI to update records, you flip one switch per server: "Allow this MCP server to modify data". Even then, Hutif adds guardrails a hurried human might forget:

  • an update or delete must name which rows it targets; "delete everything in this table" is refused,
  • at most 500 new rows per request,
  • each value is capped in size.

4. Only the right people (and apps) get through the door

Each Hutif server has its own address with a long random identifier in it. For a quick internal test, that private link may be all you need. For anything real, switch the server to Secured:

  • Sign-in with OAuth. Turn on "Require OAuth sign-in" and every AI app has to ask for permission first. You see a consent screen naming the app and the data source, and only you, the owner, can click Allow. Hutif uses OAuth 2.1 with PKCE; access tokens expire after an hour and each token works for one data source only.
  • Secret request headers. Require up to four header values that the caller must send exactly, on top of sign-in or instead of it.
  • A new link in one click. If a link was shared too widely, "Regenerate URL" makes the old one stop working immediately.

5. An off switch that works instantly

Every server has an on/off switch. Turn it off and the very next request from any AI gets nothing: no tools, no data, no cache in between. Turn it back on when you are ready.

Built-in limits cap how hard any AI can hit your data. Each server accepts a set number of requests per minute (60 on the free plan), anything above is refused on the spot, and you get an email when your account reaches 80% of its monthly requests.

6. What Hutif keeps, and what it doesn't

Security is also about what you don't collect.

  • Databases and APIs are queried live. Hutif doesn't copy them, cache them or store the answers. When your AI asks, Hutif asks your source, passes the answer back, and forgets it.
  • Uploaded spreadsheets are the exception: to serve them, Hutif keeps a copy of the rows from the file you uploaded, tied to your account.
  • Activity logs are minimal. For each request Hutif records which tool was called, when, from which address and whether it worked. Not the question, not the query, not the data that came back.
  • No AI on our side. Hutif doesn't run any AI model on your data. Your data goes to one place only: the AI app you connected yourself.
  • Your data stays yours. Every request is checked against the server's owner, so nothing from one account is ever visible through another account's server. Personal data is handled under the EU GDPR.

7. The part only you can do

Hutif controls what your AI is offered. Your database controls what the credentials you gave Hutif are allowed to do. For the strongest setup, use both:

  • create a dedicated database user just for Hutif, instead of reusing an admin or app account,
  • give that user read-only rights unless you truly need writing,
  • let it see only the tables your AI needs,
  • for APIs, use a key with the smallest set of permissions the API offers.

Do that, and you have two locks that don't depend on each other: Hutif's front desk, and your database's own permissions behind it.

Want to see every switch in action? This video walks through the six settings to turn on before your AI touches your data:

The short version

A password gives your AI a master key, the power to delete, and a secret that spreads. Hutif gives it a front desk instead: a few clear tools, read-only by default, behind a door you control, with credentials it never sees.

You can set one up in a few minutes, with no code, on the free plan: hutif.com. For the full technical details, see our security page.

← Back to blog