Let an AI-built tool use company data safely
Have an admin connect the data source once, with a login that can only read what your tools need; tools then read through Patchy Cloud and never see the password. Everyone who can open a tool can see what it reads, so connect only data your whole company may see.
Before you start
A few sources today, read-only unless you allow email. Patchy Cloud connects a short list of sources, including Postgres databases, PostHog projects and company email addresses on Pidgeon. Tools can’t write back to any of them, except that an admin can let tools send from an email address and sort its mail.
The source sees one login, not each person. Any coworker can use a connection through any company tool built on it.
Some data can’t go in at all. The beta terms rule out health information covered by HIPAA, card or bank account numbers, government ID numbers, personal information about children, personal data covered by the GDPR or UK GDPR, and data a law or contract says must be kept on certified or audited systems. Don’t connect a source that holds any of them.
What you need
- Admin access in Patchy Cloud. Only admins connect data.
- An agent set up to build Patchy tools. Start with the Claude Code or Codex guide.
- For Postgres: a database that’s reachable from the internet over an encrypted (TLS) connection, and whoever manages it, to create a login.
- For PostHog: a PostHog Cloud project. Self-hosted PostHog isn’t supported.
- For email: an address on Pidgeon, and a Pidgeon admin key.
Steps
Get a login that can only read what tools need
For Postgres, ask whoever manages your database for a login that can only read the tables your tools should see. Patchy Cloud runs every query read-only, but that promise rests on the login you give it. Logins that are too powerful, like a superuser, are refused. For PostHog, create a personal API key limited to one project, with Read access to Query, Event definitions and Property definitions, and note the project’s region and ID. For an email address, Patchy Cloud uses the admin key once to make its own key for that one address; choose Read unless tools need to send.
An admin connects it
In Patchy Cloud, open Integrations, then Connect an integration, and follow its walkthrough. Paste the connection details or key, pick the short name your agents will use, like
postgres/warehouse, and say what’s in it. Before you confirm, the form says it plainly: every member can read this through any tool that uses it.Ask your agent to use it
Name the source in your request. Your agent finds what’s connected and adds it to the tool. For Postgres, it builds from table and column names and tests with made-up rows, never your real ones. For PostHog and every other source, it reads real data through Patchy while it builds, as you, and each read shows up in the connection’s recent calls. Email the tool sends while it’s being built goes nowhere.
Say, for exampleShow open deals from postgres/warehouse in the lead tracker
Publish it
Ask your agent to publish the tool with Patchy, then open the link and check the results. Coworkers open it as themselves: before each read, Patchy Cloud checks they’re still in your company and records who it was. Your database sees only the one login you connected.
Keep an eye on it
Open a source from the Integrations page to see which live tools use it. Admins also see its latest calls there: when, from which tool, for whom and how it went. An admin can replace the login or disconnect at any time; disconnecting cuts off every tool that uses it.
Questions
Can a tool write to our database?
No. Patchy Cloud runs every database query read-only, and tools can’t change any connected source except an email address an admin lets them send from. A tool keeps its own data in its own storage inside Patchy Cloud.
Can I limit a source to one tool?
Not yet. Any company tool built on a connection can use it.
Can I see what ran?
Admins can. Each connection’s page lists its latest calls: when each ran, from which tool, for whom and how it went.
Does the agent see our real data?
While it builds, for Postgres, no: it gets table and column names and tests with made-up rows. For every other source, such as PostHog or an email address, yes: it reads real data, and admins can see each read. Separately, once an admin turns on agent access, people’s own agents can read, as them, whatever a tool’s builder lets agents read, and that can include real rows.