Figma, GitHub, your docs —wired to one agent with an arrow
Search the whole official MCP Registry right in the app, install a server in one click, and connect it to exactly the agent that needs it. The designer gets Figma. The reviewer does not.
35,000+
servers in the official MCP Registry, searched on your computer
0
keys an agent ever sees — only the tools
1 arrow
to give a server to an agent, and to take it back
The arrow is the permission
An MCP server is a card. Draw an arrow from it to an agent and that agent gets the server’s tools on its next turn — Claude Code without a restart. Agents without the arrow do not see it at all.
Figma Context
io.github.GLips/Figma-Context-MCP · v0.13.2
Connected agents: 0
Draw an arrow from this card to an agent.
2 tools figma_context__*
Every server in the official registry, found as you type
The app keeps its own index of the MCP Registry and searches it on your computer — no waiting on a server for every keystroke. Open one to see how it runs, what it asks for and exactly what will be installed.
Try it: type “figma”, pick a server, install it.
Pick one to see what it is and exactly what will run.
Install once, connect per agent
Nothing goes into your own Claude Code or Codex settings. The app runs the server itself and hands its tools to the agents you connect.
Find it
Search the official MCP Registry. Filter by servers that run locally, remote ones, or ones that need no setup.
Read the plan, then install
You see the package and exact version, the command that will run and which keys it needs. npm and PyPI packages install into NeuroSquad’s own folder; Docker images are pulled; remote servers just connect.
Add the card, draw the arrow
Put the server on the canvas as a card and arrow it to an agent. The agent sees its tools as <server>__<tool>, each described with the server’s name.
Take it back anytime
Delete the arrow and the tools disappear from that agent. The server stops by itself after a while without calls.
Honest about each harness
Servers reach agents through NeuroSquad’s own MCP connection, so it works where that connection does.
- Claude CodeRight awayRe-reads its tools mid-session — checked with a real agent.
- Codex CLIOn its next startGets the tools when it starts; restart the card after drawing the arrow.
- Qwen CodeOn its next startSame as Codex: tools arrive with the next start.
- OpenCode, Cursor, Kilo, Hermes, Pi, CrushNot supportedThe app does not connect these to its MCP server; the card says so.
Third-party code, handled like it
Everything that comes from a catalog is treated as untrusted.
You see what runs
Install shows the package, version and the exact command, and waits for your consent. Docker flags that reach your files or network are called out.
Keys stay in your keychain
Tokens are encrypted with the system keychain and handed to the server’s process only. Never in a command line, a config file or an agent’s context.
Nothing global
No global npm or pip installs, no edits to ~/.claude or ~/.codex. Uninstall removes the folder and the saved keys.
No shell tricks
Commands are built as argument lists, never shell strings; package names, versions and addresses are validated before anything runs.
Sign-in in your browser
Servers that use OAuth open their own login page in your browser; the token comes back to the app, not to the agent.
A clean environment
A server process gets the basics and its own settings — not the API keys that happen to be in your environment.
Questions
Give your agents the tools your team uses
Install NeuroSquad, open the catalog, arrow a server to an agent.