Skip to content

Use with AI Coding Agents

Vibepit works with any AI coding agent that runs in a terminal. The sandbox provides network isolation by default, and you control which domains each agent can reach through presets and allowlist entries.

Choosing a session mode

Vibepit offers two ways to run a sandbox:

  • vibepit run (default) — attaches your terminal directly. The session ends when you exit. Best for quick, interactive use.
  • vibepit up + vibepit connect — runs the sandbox in the background with an SSH server. Sessions persist across disconnects and you can reconnect with vibepit connect. Stop with vibepit down. Best for long-running agents or agents that connect via SSH.

Both modes use the same network isolation, allowlist rules, and container hardening. The examples below use vibepit run, but you can substitute vibepit up followed by vibepit connect for daemon mode.

Discovery workflow

Every agent has its own set of backend services, and those services change over time. Rather than maintaining a static domain list, use the monitor to discover exactly what your agent needs:

  1. Start a session with the default preset (included automatically on first run):

    vibepit run

  2. Launch your AI coding agent inside the sandbox.

  3. In a separate terminal, open the monitor:

    vibepit monitor

  4. Look for blocked requests. The monitor marks each request with + (allowed) or x (blocked).

  5. Allow blocked domains directly from the monitor by navigating to a blocked entry and pressing a (session only) or A (session + save to project config).

  6. Alternatively, allow domains from the command line:

    vibepit allow-http api.example.com:443

  7. Repeat until your agent operates without blocked requests.

This iterative approach means you always grant the minimum access your agent actually needs, regardless of how its backend services evolve.

Agent-specific notes

The default preset (pre-selected on first run) bundles presets for several common agents. In most cases, no extra configuration is needed beyond the discovery workflow above.

Agent Included preset Covers
Claude Code anthropic Anthropic API domains
OpenAI Codex openai OpenAI API domains
GitHub Copilot vcs-github Core GitHub domains

If you use MCP servers that fetch remote resources (common with Claude Code), enable the mcp preset via vibepit run --reconfigure and then use the discovery workflow to allow any remaining domains your servers need.

For any agent not listed here, the discovery workflow works the same way — start a session, run the monitor, and allow what gets blocked. Save entries to your project config so they persist across sessions.

For full details on allowlist management and the monitor interface, see Monitor and Allowlist.