Skip to content
vops

Development agents

A development agent can inspect a repository, prepare a flui.yaml, compare targets, propose a deployment and diagnose the result. The useful part is not giving it a shell with more credentials; it is giving it enough product knowledge and a constrained path to ask vops for semantic operations.

Human docs and agent docs stay separate

This website explains the system to people. The material an agent consumes is distributed with the vops executable:

  • a progressively loaded vops-deploy skill;
  • a short project bootstrap;
  • versioned capability and error references;
  • MCP tool schemas generated from the capability registry.

Keeping those artifacts in the product repository prevents a published website and an installed CLI from describing different tool contracts.

One setup flow, four adapters

vops has native setup adapters for:

  • OpenAI Codex;
  • Claude Code;
  • OpenCode;
  • Google Antigravity surfaces that support local tools or MCP.

The adapter changes only the files owned by that client. Setup is idempotent, backs up existing configuration before a merge and records which blocks vops manages so uninstall can preserve unrelated user settings.

Advisory mode

The first supported mode is advisory. vops enforces its own sessions, policies, approvals and semantic operations. However, an agent running as your normal operating-system user may also have a general shell and could bypass vops by calling SSH or reading files directly.

Do not describe advisory mode as a sandbox.

Protected mode

Protected mode requires launching the agent in an isolated environment without direct access to provider credentials, SSH keys or the administrative vops CLI. Its only infrastructure path would be the scoped vops interface.

That launcher and isolation layer are a later phase. They are not implied by installing the MCP configuration.

The human control loop

For controlled work, the human:

  1. connects the selected client;
  2. creates a short-lived session with an objective and limits;
  3. gives the session token to that agent only;
  4. reviews approval requests and operation state in the CLI or local dashboard;
  5. pauses, revokes or stops all sessions when needed.

The agent cannot approve its own request through the MCP tool surface.

Start with Choose an agent approach before moving to Client setup.