Agents

An agent is the coding-assistant runtime a session wraps. Consortium doesn’t reimplement agents — it wraps the CLIs you already use, adds device-portable sessions, and syncs everything end-to-end encrypted.

First-class agents

Command	Agent	Runs against
consortium claude	Claude Code	Your Anthropic subscription/key, or managed credits
consortium codex	Codex	Your OpenAI subscription/key
consortium gemini	Gemini CLI	Your Google account/key
consortium grok	Grok	Your xAI account/key
consortium pi	Pi	Anthropic models
consortium code	Consortium Code (opencode)	Free managed models by default; managed catalog with credits

Each launches in the current directory and registers the session with your account. Per-agent model options are listed in Supported models.

Choosing

• claude is the most mature integration — permission prompts, tool approval from the app, and Project Manager mode all target it first.
• code (Consortium Code) is the zero-setup path: it runs on the free managed models with no vendor account.
• codex, gemini, grok are the right choice when you’re standardized on that vendor’s stack or billing.
• pi favors a stricter permission posture over raw capability.

Orchestration

Two platform features drive agents for you:

• Automations — run an agent with a prompt on a schedule; each run becomes a normal session you can open and steer.
• Project Manager mode — a Claude Code session started in PM mode gets extra tools to list and inspect your other sessions, send them prompts, spawn sessions (including in git worktrees), manage project cards, and open pull requests. It turns one session into a coordinator for the rest.

Custom agents

A declarative adapter for registering your own agent runtime is in development. Today, adding an agent means wiring its spawn, RPC, and session-replay paths in the CLI — feasible (all six above go through that path) but not yet a supported extension point.