67 coding agents, one harness

Shuriken runs 67 coding agents in one app: Claude Code, Codex, GitHub Copilot, OpenCode, Gemini CLI, Antigravity, Cursor, Grok, Amp, goose, Cline, Aider and many more. Each chat picks its agent. One project can hold a Claude Code chat, a Codex chat and a Gemini CLI chat side by side, each on its own branch, all on your own computers.

Why we built it this way#

We kept hearing the same thing from people who pay for more than one AI plan: "I use Claude Code for this, Codex for that, and Gemini when the other two are out. Every one of them is a separate terminal, a separate config, a separate set of sign-ins."

So we didn't build an agent. We built a harness: the part around the agent. The agent does the coding. Shuriken gives every agent the same project, the same chat list, the same MCP servers, the same queue and the same phone view. How a harness differs from an orchestrator or an IDE is in what an agent harness is.

All 67#

The list comes straight from the catalogue in our code, so the number on this page is the number in the app. Agents get added as new CLIs appear.

Your sign-ins, not ours#

Every agent runs on accounts you already have. Add each subscription sign-in or API key once, as a source: Claude, ChatGPT for Codex, Gemini, Copilot, Cursor, Grok, Kimi, Qwen and more, plus your own endpoints. Several per provider is fine. Sources sync to all your devices, sealed end to end.

Shuriken never bills model usage. Your Claude plan pays for Claude Code, your ChatGPT plan pays for Codex, and so on. The Free plan holds two sources, so one Claude and one ChatGPT works out of the box.

How deep each agent goes#

We'd rather tell you this up front than have you find out.

Claude Code gets the most, because it gives Shuriken the richest output to work with. Its chats show subagents, workflow phases, background commands and monitors live in Activity. When its account reaches your switch point, a Claude Code chat moves to your next Claude account on its next request, mid-task, with its subagents.

Codex, Copilot, OpenCode, Pi and Oh My Pi are built in as full Shuriken chats too.

44 more open as a Shuriken chat through the Agent Client Protocol: Gemini CLI, Antigravity, Cursor, Grok, Amp, goose, Cline, Kiro, Qwen Code and most of the rest of the list. Antigravity chats run on Google's own ACP server, which Shuriken installs with the Antigravity CLI. You get the chat, the queue, the worktree and the phone view. When an account runs low, they move to another account when their current turn ends.

17 have no chat protocol, so they run in their own terminal interface inside Shuriken: Aider, Continue, Codebuff and others. Still on their own branch, still on your sources, still reachable from your phone; just less structure for Shuriken to show.

Agents that have a terminal interface of their own, Claude Code and Codex included, can also be opened in it from a chat tab (Open in terminal) when you want the native experience.

One project, any mix#

Every chat except the Main chat gets its own branch in its own git worktree, so a Claude Code chat and a Codex chat never edit the same files under each other, and there is no cap on how many run at once (parallel agents has the details). Add an MCP server once in Global MCPs and every Claude Code, Codex, Gemini CLI, Copilot, Cursor and OpenCode chat is handed the same servers when it starts, and you sign in to each server once.

To compare two agents on one task, give the same prompt to two chats, compare the diffs in Changes, and finish the one you like. The git commands behind it, and how to avoid conflicts and port clashes, are in our guide to running agents in parallel.

Each chat switches between the sources you added for its agent. Shuriken never moves a Claude Code chat into Codex: that's your call, made by starting the next chat with the other agent.

Missing CLIs get installed for you#

The model picker lists the agents a device can run first. One that isn't installed shows how to install it, and some are fetched automatically the first time you use them.

On a new computer, setup installs missing tools for you, and Check again finds ones you installed yourself. On a server (Linux, macOS or Windows, added from Settings › Devices), the installer sets up the agent CLIs and GitHub sign-in and tells you which step it's on. So "I'll set up Codex on the build box later" stops being a task.

Which one should I use?#

We're not going to tell you. It changes with every model release and depends on your code. Our Claude Code vs Codex comparison lays out the differences from each vendor's docs, without a winner. Claude Code also has its own pages against Cursor, OpenCode and Amp, and Claude Code alternatives covers the rest. The fastest way to find out is to run the same task in two chats.

Every agent is on every plan, Free included. Download Shuriken and add your first two sources.