Agents that steer Offsite

You tell an agent what to start, and it starts it: a Claude session can open a Codex card, a Codex session a Grok card. You say it in the session you are in, typed or by voice, on the Mac or on the iPhone. The agent uses Offsite’s tools on the host, and the board follows on every device.

0:43

One sentence on the iPhone, and Claude starts Codex with the model you named; Codex gets its brief from Claude while you two carry on. A 43-second film, with sound.

Start another agent

“Start a Codex session with Astra and have it fix the flaky cart test, while we focus on the checkout.”

Claude opens a new card next to its own, with Codex in it and the model Astra selected. Codex’s first message is a brief that Claude writes for it. Claude tells you which card it started and goes on with you on the checkout. Neither session waits for the other.

Fresh or forked

A new session starts one of two ways.

Start another agent

1. Freshknows only its tasktask→new card
2. Forkedreads where this one standslast 300 entriestask→new card

A fresh session starts from its task alone; a forked one first reads the last 300 entries of this session's transcript.

Fresh: the new agent gets its task and nothing else, so the agent that starts it writes the brief to stand on its own.

“Have Grok write the tests for the cart while you keep going.”

A Grok card opens with the tests as its task. It knows only that task.

Forked: the new session first reads where this one stands, the last 300 entries of its transcript, then starts on the task.

“Start a new session from here and build the receipt emails, while we go on with the checkout.”

A second card opens, knows the work so far and builds the receipt emails, while the first carries on with you. It is the same move as Hand off to a new session… on the card’s menu (Hand off and ask), started by the agent. Nothing is written into the running session.

The new card

• Agent: the one you name, Claude, Codex or Grok. Without a name, the same agent as the session that starts it.
• Model and effort: the ones you name. The agent passes the model as the CLI names it (for Codex gpt-6-astra, for Claude also opus or sonnet), and Offsite hands it to that CLI unchanged (--model, for Codex -m). The CLI checks the name and answers an unknown one in the new card. Without one, the CLI’s default.
• Name: the card starts under the name the asking agent gave it, or the agent’s name, and renames itself once it knows its work.
• Where it lands: to the right of the card that asked, in the same workspace. A session started in another workspace goes to the left edge of that board.
• Permission mode: the same agent inherits the mode of the session that started it, so a YOLO Claude starts a YOLO Claude. A different agent starts in its own default mode.

More you can say

You say	What happens
“Fix this bug, commit and push, and close the session.”	The agent does the work, writes a last recap, and when its turn ends the card moves to the Archive.
“Carry on in a fresh session, and close this one.”	A forked session continues in a fresh context; this card moves to the Archive.
“Check what Codex found and have Grok fix it.”	The agent finds the Codex session, reads its transcript, and starts a Grok session with the findings in its brief.
“Start a Claude in bloom-admin for the API client.”	A Claude card opens in the workspace bloom-admin, with the API client as its task.
“What is the session Fix login doing?”	The agent reads that session’s transcript and answers here.

Read other sessions, across tools

The agent can list your running agent sessions on the host, with card name, agent, workspace, whether each is busy or idle, and its latest recap. It reads a session’s transcript by its card name or ID, whichever agent runs there: Claude reads Codex, Codex reads Grok.

This is a read from disk. The other session is not contacted, not interrupted, and the agent cannot send anything to it. To pass work on, it starts a new session with what it read in the brief.

Close the session

Tell the agent to close its session, on its own or after a task. It writes a last recap line, and when its turn ends the host moves the card to the Archive. The conversation stays, and you bring the card back from there like any parked session (The Icebox covers both park lists).

What asks you first

Reading other sessions, naming its card, updating its recap and closing itself: for Claude Code and Codex these never ask. Grok applies its own approval to every tool.

Starting a session starts work that runs on your subscription, so it goes through the agent’s own approval, like a command that does real work. In YOLO the agent goes ahead. In Auto the CLI’s approval rules decide whether it asks you first. See Permission mode.

From any device

The tools run on the host, not in an app. Say it to an agent on the iPhone while you are out, and the new card shows up on the Mac’s board too, and the other way round.

Details

• The tools sit in the offsite_computer MCP server that Offsite gives every Claude, Codex and Grok session: session_start, agents_list, agents_context and session_finish. What each one takes and returns is in In daily use.
• Starting and closing a session are written to ~/.offsite/agent-bridge/audit.log on the host, with the session that asked.