Supported agents and what each can do

You want your agent picked back up at 03:00. Can Agent Auto-Continue reach it? It works with six: T3 Code, OpenCode, Claude Code, Codex CLI, Claude Desktop and Codex in the ChatGPT app. Every one can be scheduled for a time and run several turns with a stop phrase. Two things matter most. Can the app see the usage limit, so When free works? And can you lock the Mac while it waits?

The short version

Agent	Sends through	When free	Reset time	Screen locked	Needs running
T3 Code	Local API	No	No	Works	T3 Code
OpenCode	Local API	Partly	Partly	Works	opencode serve
Claude Code CLI	Headless mode	Yes, inferred	If stated	Works	Nothing
Codex CLI	App-server protocol	Yes	Yes	Works	Nothing
Claude Desktop, Code sessions	Accessibility	Yes, inferred	No	Waits for unlock	The app
ChatGPT app, Codex threads	Accessibility	Yes	Yes	Waits for unlock	The app

When free sends once the app sees the limit lift; where it can’t see limits, pick a time. Inferred means read from local records, and an inferred limit with no reset time doesn’t hold a send back.

Agent by agent

• T3 Code. Needs a bearer token (install has the command). Checked against Alpha 0.0.40.
• OpenCode. Run opencode serve and set its port in Settings > Agents. The app only sees a limit a session’s already waiting on, with OpenCode’s next retry time. Otherwise When free sends straight away. Checked against 1.18.34.
• Claude Code CLI. It runs headless claude -p --resume, so exit the session after scheduling (while it’s open, the message waits). Limits are inferred from Claude Desktop’s usage samples or recent transcripts. It reuses the session’s permission mode; headless runs deny anything that would still prompt and carry on. Checked against 2.1.286; see the Claude Code page.
• Codex CLI. It speaks the app-server protocol, through the shared daemon if one’s running, a private server if not. Approval requests get the answer that stops the turn, and nothing more is sent. Quitting the app interrupts private-server turns, not daemon ones. Checked against codex-cli 0.159; see the Codex page.
• Claude Desktop. The app fills in its message box and presses send through macOS Accessibility. That needs the app running, the Accessibility permission and the Mac unlocked. With no reset time, When free can hit a limit; that try doesn’t count and the next waits 15 minutes.
• ChatGPT app. Same method, same three needs.

Codex threads stay with the app that made them, so CLI, ChatGPT app and T3 Code threads each carry their own agent’s mark. Claude Desktop’s sessions are likewise hidden from the Claude Code list.

Choose which agents show, and in what order, in Settings > Agents. Hidden ones aren’t checked.

If it can’t send yet

A one-off message due while the agent’s busy, the Mac’s locked or the agent’s unreachable doesn’t fail. It waits and checks again, for up to six hours. An agent waiting on your answer to a question or permission prompt gets nothing, since only you can answer. The desktop apps need the Mac left unlocked, and keep-awake can’t unlock it.

The app checks Claude Desktop and the ChatGPT app read-only at launch, when something’s scheduled for them, and every five minutes while that’s pending. If an update to either changed what the app relies on, the dashboard flags it before the schedule’s due.

What it can’t reach

• ChatGPT and Claude Desktop chats. They live in the cloud, with no local record to find the conversation or confirm a message arrived. Matching titles on screen would be too flaky to send unattended.
• Codex threads from an IDE extension. Neither Codex agent claims them.
• Knowing the task’s done. No agent reports that, so the turn limit and Stop are the real guards.