MCP servers, set up once for chosen projects

You can now set up an MCP server once and say exactly which projects get it. In Settings › Extensions › MCP servers, each server runs in All projects or in the projects you pick. Your agents' chats in those projects get the server when they start, and one sign-in covers all of them. It's part of Shuriken 5.1.

Why we built it#

Global MCPs solved the first problem: signing in to the same server a hundred times. You sign in to GitHub's or Linear's MCP server once, and it works on any account and any device, because the sign-in belongs to the server, not to your AI account.

Then we kept hearing about the second problem: the database server showing up in a blog repo, the design tool's server sitting in a backend project. Global meant everywhere, and everywhere is too much.

Every MCP server adds tools the agent can see, and depending on the agent, their descriptions take room in its context. A project that never touches Figma has no use for Figma's tools. Fewer servers means less clutter and fewer wrong tools for the agent to reach for.

So we kept "set up once" and added "only where you need it".

How it works#

Open Settings › Extensions › MCP servers. It lists every server you use, once, each with where it runs.

To add a server, pick one of the 57 ready-made connectors (GitHub, Linear, Notion, Slack, Sentry, Supabase, Playwright and more) or add any server by address or command. Shuriken asks one question: Which projects can use it? Leave All projects on, or turn it off and switch on the projects you want.

To change your mind later, open a server to see where it runs. Add a project, remove one, or switch back to All projects. Agents in a project pick up the change when they next start.

If the server needs a sign-in, you do it once. Every project that uses the server shares it, on every one of your computers and servers. Sign-ins travel between your devices sealed in your vault, and our cloud keeps a key that opens the vault.

What Shuriken writes for you#

You don't edit config files. A server on All projects goes in Claude Code's user config on that computer, so every project there loads it. A server for chosen projects goes in that project's section of Claude Code's config on that computer, the same place claude mcp add puts one by default. It lives outside the project folder, so it never shows up in a chat's diff or lands in a commit.

Your own .mcp.json stays yours. If a project already has one, Shuriken lists its servers and can switch them on and off, but it never creates the file for you.

Every agent gets the same list. When a chat starts, Shuriken hands it the servers its project uses, whether the agent is Claude Code, Codex, Gemini CLI or another one, and whichever of your accounts it runs on. You set it up once, and all of them see it.

One honest exception: servers you added straight into Codex's own config file are global to Codex. Codex has no per-project switch for those, so turning one off turns it off everywhere, and Shuriken's list says so on that row rather than pretending otherwise.

An example setup#

A split for someone with a few repos might look like this.

That last row is the one people ask about most: add the same connector twice, give each its own sign-in, and assign them to different projects (Two Notion workspaces, one MCP list walks through it).

Good to know#

When Shuriken moves a chat to another of your accounts at a limit, its MCP servers and their sign-ins stay put. Each device runs what it can: if a server needs something a computer doesn't have, its row says so, with the fix. And it's on every plan: Global MCPs, chosen projects included, are part of Free.

New to MCP servers? Start with Set up MCP servers once for every project, then pick a few from Best MCP servers for Claude Code. To see it for yourself, try Shuriken and open Settings › Extensions › MCP servers.