Claude Code, Codex and opencode start in a dock inside SnoutData, in your workspace, with your connections handed over the moment they open. No .env file, no connection string in a repo, no copy of your password.
A real recording. The agent asks, you allow, and the app moves.
An agent gets an icon in the activity bar when it is installed on this computer and you have not switched it off.
Anthropic's CLI, handed the local broker by flag, so it has your databases from the moment it starts.
OpenAI's CLI, in the same dock, in the same workspace folder, with the same access.
The open-source alternative, treated exactly the same way as the other two.
Closing the dock detaches the session rather than killing it, so a long job survives you putting the panel away.
The app publishes what it can do as tools, so the agent works the product rather than describing it.
Through the app, over the connections you allow, read-only by default.
Opens tabs, puts SQL in the editor, changes settings, reads app state and logs, posts notifications. The tools run in the live interface, so the window updates while the agent works.
Anything it has no tool for it can still show: markdown comes back as prose, a SQL block as a card with a Run button, a chart as a chart. The console has no message box, because the terminal already is one.
Ask for a dashboard and it opens the view and creates the widgets, instead of printing SQL for you to paste.
The agent has to ask, and the asking is the point.
Your passwords and keys stay in your operating system's keychain. The app hands the agent capabilities, not secrets, so nothing it is given contains your database password.
Anything that changes the app is denied until you say yes. When it is an action you will allow again and again, you can grant it standing permission by name, instead of clicking through the same prompt forever.
An activity timeline in the Agent log, a notifications centre, an entry in the same AI audit trail the assistant writes to, and an indicator in the status bar while it is acting.
Agent access is opt-in per connection and off by default, so the agent sees the databases you picked and no others.
The same broker, for any MCP client.
If you would rather keep your agent where it already lives, SnoutData runs a local MCP server on your own machine, behind a bearer token, offering the same read-only tools over the connections you opt in. Your agent gets the data, your credentials stay in the keychain, and every call it makes is written to the same audit log as everything else.
Free to start, on Windows, macOS and Linux.