Skip to main content
Agents are only as good as the data they can work with. Mocked data misses edge cases. Shared staging gets corrupted. Production is off-limits. Ardent gives every agent run its own isolated copy of your real database: spun up in under 6 seconds, handed off as a connection URL, thrown away when the task is done. There are two ways to run it: you create the branch and hand the agent a URL, or the agent manages branches itself.

Drive it yourself: hand the agent a URL

Your script or CI harness runs the CLI; the agent only ever sees a branch URL. The agent needs no Ardent token and can’t touch anything but its branch.
Nothing reads AGENT_DATABASE_URL automatically — pass the URL in the agent’s prompt, or export it as DATABASE_URL if your app reads that.

Let the agent drive: set it up once

The agent runs the Ardent CLI itself — creating, using, and deleting its own branches:
For a single task, paste this into Claude Code or Cursor:
If your agents touch the database every day, install this as a skill instead of pasting per task. Install into your project (or into ~/.claude/skills/ to use it everywhere) and Claude Code loads it automatically whenever a task involves Ardent:
The skill keeps itself current: each time it runs, it fetches the latest version of itself from these docs and follows that copy. If the fetch fails (offline, blocked network), it falls back to the directions already installed.
To pin the installed copy and skip the self-update, set SKIP_ARDENT_SKILL_SYNC=1 in the shell that launches your agent (for Claude Code: export it before running claude, or set it in .claude/settings.json under env). A Python venv does not carry it.
Skills load when the task looks relevant — so a request that never mentions the database might skip it. To guarantee it’s used, add this to your project’s CLAUDE.md, which Claude Code reads every session:
This is what it installs:
The skill deliberately contains no command arguments. It tells the agent to fetch the current reference from these docs before acting, so when the CLI or API changes, the docs update and every installed copy of the skill stays correct without being touched.
Real data, isolated per run, no production access.