Why we call Shuriken a harness

Since 21 October 2026, we describe Shuriken as a harness for coding agents. The word we used before came from the concert hall. It's one word, but the new one is right, and the old one was quietly promising something we don't do.

What a harness is#

A harness runs the agents you already use, with your accounts, your tools and your devices. The agent (Claude Code, Codex, Gemini CLI and so on) does the thinking and the coding. The harness does everything around it: which account it runs on, which MCP servers it can reach, which computer it runs on, and how you follow it from wherever you are.

A harness makes good agents easier to live with. It doesn't try to be a smarter agent. (The guide what is an agent harness has the full definition.)

Why we changed the word#

The concert-hall word suggests a conductor: something that takes your goal, breaks it into a plan, hands the pieces to a crew of agents and merges what comes back. Some tools do that. That isn't what people use Shuriken for.

What we kept hearing from the people using it was much more practical:

• an account ran out mid-task, and carrying on meant logging out and in again;
• the same MCP server wanted a fresh sign-in every time they switched account or project;
• the laptop went to sleep and took the agent with it;
• they wanted to check on a chat from their phone.

None of those are orchestration problems. They're harness problems. So we fixed them, and then we noticed our own homepage was describing a different product.

What Shuriken does as a harness#

Shuriken runs 67 coding agents: Claude Code, Codex, GitHub Copilot, Gemini CLI, Cursor, OpenCode, Amp, goose and many more, each with the sign-ins you already have. You can run as many in parallel as your machine and plans allow, and any chat can work on its own branch in its own worktree if you pick that when you start it.

You add the accounts and API keys you own and are entitled to use. When the one in use reaches its limit, source switching moves your chats to the one with the most room and they keep going. 5-hour clocks start idle Claude accounts' 5-hour windows with one tiny background request each, so limits come back early.

There is one list of MCP servers for every project and account, and you sign in once: with Global MCPs the sign-in belongs to the server, not to your AI account.

The agents run on your own computers and servers. Our cloud links your devices and relays end-to-end encrypted messages. It runs no agents and stores no code. You can close the window, close the lid with Cocaine mode (yes, that's its name, and yes, it keeps your Mac awake with the lid shut), and follow everything from your phone in the web app.

And one thing it never does: bill you for model usage. Your agents use your plans and your keys.

About our own orchestration mode#

Honesty corner. Shuriken does have an orchestration mode, where a lead agent plans and workers build. You'll see it in a project's settings, marked WIP, and you can't turn it on yet. It's an experiment. It's not ready, and we'd rather tell you that than market it.

Maybe it grows up one day. If it does, it'll be a feature of the harness, not the description of the product.

What changes for you#

Nothing in how Shuriken works. You'll see "harness" across the app and the site, and the old word disappearing from both over the next releases. Agents are still agents (Claude Code, Codex…), providers are the companies behind them, and sources are the accounts and keys you add.

The longer story of why Shuriken exists at all, which starts with logging in and out of Claude, is in why we built Shuriken. To see the harness at work, try Shuriken.