co ai Delegates to Codex and Claude Code — and You Watch It Work

co ai has always run a coding agent against your project. In 1.7 it can hand a task straight to Codex or Claude Code and stay honest about what is happening. The Work Room shows current state: what the delegated agent is doing right now, what it wants approved, what it just finished. Not a spinner that hides everything, and not a raw terminal transcript dumped into chat — a summary you can trust, with the detail one click away.

Close the tab and come back: the session is still there. The Work Room reconnects to the same live work whether you left through a browser refresh, a dropped connection, or a closed laptop lid.

Three permission modes, plain names

1.7 collapses the permission system into three modes anyone can hold in their head:

• read-only — every tool call asks a human first; reads pass.
• auto — the default. Reversible workspace work just runs; anything external, destructive, or credential-touching stops and asks.
• full-access — nothing asks, bounded by an explicit turn budget you set. When the budget runs out, the session drops back to auto. A bounded grant, never a standing state.

The server owns this state — a client cannot talk itself into more authority — and the same three words appear everywhere: the CLI, the browser, the Control Center.

Activity that reads as intent

Tool cards now lead with a short natural-language line — what is being done and why — written by the model that made the call. The SQL string, the shell command, the raw result: all still there, collapsed one click away. Watching an agent work should feel like reading a colleague's status updates, not tailing a log file.

Terminal or browser, same agent

Run co ai in a terminal for a direct session, or let it open a web chat at chat.openonion.ai — both talk to the same coding agent over the same authenticated connection. Todos update live, approvals are explicit, and a hard stop stops the actual running work.

Upgrade

pip install --upgrade connectonion

See the co ai guide and the permissions guide. The 1.7 line now takes stabilisation patches; new feature work — remote browser sessions and async execution — is the 1.8 preview train.