Let Claude Code call Codex, Gemini and other agents

In Shuriken 5.3 the agent in any chat can start background agents and chats on your other providers, for any task, with hot swapping in each.

FeaturesBy ShurikenUpdated 7 November 20264 min read

Contents

From Shuriken 5.3, the agent in any chat can start work on another provider. Claude Code can call Codex, Codex can call Gemini CLI, and any of them can start Cursor, Copilot, OpenCode, Kimi or another agent Shuriken runs on that computer. What the work is for is up to you: ask in plain words, and the agent picks the tool.

What you can ask for#

Background agents. Up to 8 in one call, on the providers and models the agent picks or you name, each with the same task or its own. They are listed under the chat that started them, and that chat hears when each one finishes. By default they work in the chat's folder, the way its own subagents do. When two of them would edit the same files at once, or you may want only one's work, the agent puts each on its own branch instead.

A new chat. The agent opens it in one of your projects with a first message, on any provider, and it shows up in the sidebar like a chat you started yourself.

The same chat, carried on by another provider. The agent writes a handover (the goal, what was decided and done, what is left, the files involved), and a new chat on the provider you picked starts from it. The new chat links back to the old one, and the old one stays as it was.

Some requests it can act on:

"Have Codex write the tests for this while you do the UI."
"Ask Gemini to review this diff and tell me what it would change."
"Run the migration on Codex in the background; I want to keep talking here."
"Claude is at its weekly limit. Carry this chat on in Codex."
"Try this on Claude, Codex and Gemini, each on its own branch, and merge the one whose tests pass."

How Claude Code calls Codex in Shuriken#

Shuriken gives every agent a small set of tools for this. You never type the names; you can watch the agent use them in the chat.

Tool	What it does
`list_providers`	What each of your computers can run: installed, can start now, models, modes, usage left, resets
`start_agents`	Background agents under this chat, in its folder or each on its own branch
`read_result`	One of those agents' final answer, and its branch, commits and files changed when it has a branch
`new_chat`	A chat in one of your projects, on any provider that can start
`continue_chat`	A new chat on another provider that carries this one on from a handover
`adopt_result`	Optional: merges an agent's branch into the branch it was cut from, with Finish

adopt_result is there for the times you want one agent's branch and not the others'. It runs Finish: commit what is left, rebase, run the checks, squash-merge the branch, then archive that agent's chat and delete its branch.

Your accounts and usage left#

Before it starts anything, the agent can see which of your accounts each provider would use and how much of each is left: the share of its 5-hour and weekly limits used, when each resets, and the provider's total in fuel. So "use whichever has room" is a request it can act on.

Every agent it starts is an ordinary Shuriken chat on one of your plans. It counts toward that plan's limits like any other turn, and hot swapping (Shuriken moving a chat to another of your accounts when one nears its limit) works in it as usual. When every account of a provider is at its limit, the agent still starts there and waits for the reset, as a chat you opened would, and the chat that started it hears once that it is waiting.

The agent never gets a key, a sign-in, an e-mail address or a cost. Each account goes by the name you gave it in Sources, or by its plan when that name is an e-mail address.

What stays with you#

Only a chat you are working in can start chats or agents. Both switches are in Settings › Agents: Shuriken in every chat, and under it, Let agents ask to act. Turn the second off and no agent starts anything for you.

One chat can start at most 24 chats an hour, counting all three kinds, so a confused agent can't fill your sidebar. An agent reads the results only of agents its own chat started: another chat's messages stay out of its reach. The agents it starts cannot start more.

How this differs from running agents side by side#

Parallel agents is you opening several chats and giving each a task. Here the agent in your chat does the opening and the waiting, on whichever providers the work calls for, and tells you what came back. Using Claude Code and Codex on one project by hand is covered in Claude Code and Codex in one project.

Agents on any provider come with Shuriken 5.3, on every plan. Download Shuriken and add the agents you already use.

Frequently asked questions

Which providers can an agent start?

The ones installed on the computer the project runs on, with a source in Shuriken. The agent checks with list_providers first. A provider that isn't installed, isn't signed in or has no source is refused, and the refusal names the ones that can start.

Are these Claude Code's own subagents?

No. Claude Code's subagents run inside its own session, on Claude. These are Shuriken chats, each on the provider the agent picked, listed under the chat that started them, with hot swapping like any chat you open.

Do the agents get their own branch?

Only when the agent asks for one. By default they work in the chat's folder, like its own subagents. The agent gives each its own branch when they would edit the same files at once, or when you may want to keep only one's work.

Can the agent see my sign-ins or keys?

No. list_providers gives each source's name, plan and how much of its 5-hour and weekly limits is used, with the reset times. Never a key, a sign-in, an e-mail address or a cost.

How do I turn it off?

Settings › Agents › Let agents ask to act. With it off, no agent starts a chat or an agent for you. One chat can start at most 24 chats an hour, and up to 8 agents in one call.