Agents and models

Two words carry this page:

An agent is a coding assistant program that reads, edits, and runs code. OpenCode, Claude Code, and Gemini are agents.
A model is the AI model an agent uses to think. One agent can run many models.

Admiral is the control station. It is not an agent itself.

The agents

Agent	Status	Notes
OpenCode	Primary	The default agent. Admiral provides a working copy when your machine has none.
Claude Code	Native, first class	Sign in with your Claude account. Claude models selected in Admiral run here.
Gemini	Native, first class	Runs on Antigravity, shown as Gemini. Sign in on each device.
Codex, Cursor, Grok, Hermes, Pi, Devin	Legacy	Behind a switch in Settings. Their old sessions stay readable.

Routing

When you pick a model, Admiral routes it to the agent that runs it best:

Claude models run on Claude Code.
Gemini models run on Antigravity, displayed as Gemini.
Everything else runs on OpenCode.

A Claude or GPT model that you reach through OpenCode is an OpenCode session. It is not a Claude Code session, and the tools, sign-ins, and behavior can differ.

Switching

Two switching behaviors exist, and they are deliberately different:

Same agent, new model: the session resumes with the new model. Nothing is lost.
New agent: the agent starts a fresh native session. Admiral hands it a compact summary of verified project facts. The visible transcript stays, but the new agent has no access to the old agent's hidden reasoning or internal tool state.

Runtime management

Admiral manages the agent runtimes so you do not:

It starts and shares agent processes per project folder.
It provides OpenCode when the machine has no supported release, with setup status in Settings.
It recovers a run after an engine restart instead of losing it.

Sign-in, enablement, and usage meters for all of this live on the Providers settings page. See Settings pages.

Next steps

Switch agents and models: the task guide.
Where your data lives: what the agents may touch.