DataTugPull more from your data
Open app
☰

For your role · People wiring Claude, Cursor or Codex to company data

For AI agent builders: data access with limits

An AI agent needs real data to be useful, and a database login gives it far more than the task. DataTug Perimeter is designed to put people and agents behind the same rules: scoped reads, sensitive fields masked, changes approved. It is in development; local access policies work today.

See DataTug Perimeter All roles
Watercolor of a harbour keeper handing an apprentice a key ring with exactly one key, a closed key cabinet behind them and a tugboat moored nearby.

The agent is useful only if it can read our data. I cannot give it a login that can read everything and change anything.

Your day

  • All or nothing. The database offers a login that reads everything, or no access at all. The task needed three tables.
  • Secrets in prompts. Connection strings end up in configs, prompts and chat logs that nobody reviews.
  • Writes you did not intend. An agent that can read can often write, and “just fix that row” is one tool call away.
  • No record of what it read. Afterwards you cannot say which rows the agent saw or which customer emails reached the model.

One task, before and after: let a coding agent investigate one support ticket

Who: People wiring Claude, Cursor or Codex to company data

Today

3 tools0 hand-offs1 shared password

  1. The agent · Agent configPaste a read-write connection string into the agent’s config.a password changes hands
  2. The agent · DatabaseHope it reads only what the ticket needs.
  3. The agent · DatabaseLet it run its own UPDATE when it finds the cause.
  4. You, afterwards · Chat logReconstruct what happened from the chat log.
Today3 tools · 0 hand-offs · 1 shared passwordPaste a read-write connection string into the agent’s config.Hope it reads only what the ticket needs.Let it run its own UPDATE when it finds the cause.Click to see

With DataTug

1 workspace1 hand-offno shared password

Some steps are planned or in development.

  1. Policy ownerLimit which sources, tables and fields the agent may read, with local access policies today.Access policiesAvailable
  2. The agentPut the agent behind one Perimeter, never allowed more than the person it acts for (in development).DataTug PerimeterIn development
  3. The serverHave sensitive fields masked on the server before a value reaches the model (planned).Meaning-driven masking on the serverPlanned
  4. One person approvesTurn a change into a request a person approves, one narrow action at a time (planned).Approve an investigation stepPlanned
With DataTug1 workspace · 1 hand-off · no shared passwordLimit which sources, tables and fields the agent may read, with local access policies today.Put the agent behind one Perimeter, never allowed more than the person it acts for (in development).Have sensitive fields masked on the server before a value reaches the model (planned).Click to see

Each step names the part of DataTug it relies on and that part’s status today. “Planned” and “In development” steps describe where the product is going, not what you can use now.

Today’s access policies are enforced by your local DataTug CLI: an agent that can read your machine’s DataTug configuration could get past them. Server-enforced access through one Perimeter is in development. How access policies work

Solutions for you

Also useful: Incident Investigation

Read more: AI agents with ungoverned database access

Give your agent only what it needs

The Perimeter page on datatug.ai has the copy-and-paste install prompt for your agent.

See DataTug Perimeter