Run an agent¶

After setup, use the launcher from the ExOShell checkout. These examples assume a selected policy and the appropriate provider instances in your configuration.

./run-exoshell-agent.sh /path/to/project
./run-exoshell-agent.sh --agent claude /path/to/project
./run-exoshell-agent.sh --agent opencode /path/to/project

Codex is the default. The project defaults to the caller's current directory. Each launch creates a separate sandbox; leaving the agent deletes it by default.

Pass agent arguments¶

Options before -- belong to ExOShell; everything after it goes unchanged to the selected agent:

./run-exoshell-agent.sh --agent codex /path/to/project -- exec 'Explain this project'
./run-exoshell-agent.sh --agent claude /path/to/project -- -p 'Explain this project'

Use ExOShell's --model and --effort options for starting selections, including OpenCode's interactive TUI. Supported values depend on the agent and backend. See models and effort.

Choose what to share¶

By default, only the selected project is mounted at /workspace. Use --host-share for a larger source tree, or --no-share for a disposable workspace. See projects and workspaces.

Git commits use the selected project's effective host Git name and email when available. Configure these on the host for the project before launching. Forge host selection and provider credentials are separate; see GitHub and GitLab.

Diagnose a launch¶

./run-exoshell-agent.sh --verbose /path/to/project

Verbose output shows the selected config, effective settings, and resolved paths. Use --keep to retain the sandbox after exit for investigation, then delete it when finished. See sessions and troubleshooting.