Connect your agent¶

Run existing Agent CLIs, scripts, and automation through one entry point, without changing the agent. Install the command in the selected executor and configure the provider in the agent itself first. An agent installed on the host does not automatically appear in an image. Replace <agent-command> below with the actual command:

pvisor run --safe -- <agent-command>
pvisor status --review last
pvisor apply last --path src

--safe mode matches the executable filename and allows model API destinations on HTTPS port 443; other destinations are denied. Deliver credentials explicitly with --pass-env NAME or let the Gateway hold them. HOME writes are discarded after execution, so do not depend on them to persist login state. See credentials and environment for delivery methods.

Agent	Executable	`--safe` default allowlist	Guide
Claude Code	`claude`	`api.anthropic.com`	Claude Code
Codex CLI	`codex`	`api.openai.com`, `chatgpt.com`, `ab.chatgpt.com`	Codex CLI
Gemini CLI	`gemini`	`generativelanguage.googleapis.com`	Gemini CLI
ZCode	`zcode`	`api.z.ai`, `open.bigmodel.cn`	See `--safe` presets
aider	—	No preset; needs `--overlaynet-allow`	aider
OpenCode	—	No preset; needs `--overlaynet-allow`	OpenCode
Any script	Any	None; `bash`, `sh`, `zsh`, `fish` reuse the Codex destinations	Any script

Unknown commands deny outbound access. --overlaynet-allow HOST:PORT grants it explicitly and replaces preset lists. Matching uses the direct executable filename; shell wrappers do not trigger agent-specific adaptation.

After a run, check the installed controls and observations in the Run Bundle. Supported agent version ranges are not systematically tested; see end-to-end tasks.

Experimental integrations

The aider, Gemini CLI, and OpenCode guides use only existing pVisor options and stay experimental until a pinned-version regression exists. The Linux host selective proxy is cooperative; use a prepared VM when you need enforcement.