Ergod works both ways. It connects out to your MCP servers and uses their tools inside a real agent loop — and it serves as an MCP server itself, so Claude Code, Cursor or VS Code can call into your hosted sandbox and get real files back. Nothing to install, and no local runtime that has to stay awake.
MCP — the Model Context Protocol — is an open standard for giving an AI assistant tools. A server exposes tools: read this database, search these docs, open a pull request. A client is the thing holding the model, which connects to those servers and decides when to call them. Claude Desktop is a client. Cursor is a client. Ergod is a client too, and unlike those it is hosted, so the connection does not depend on your laptop being awake — and it is a server as well, which is the section three cards down.
A local MCP client is only connected while the app is open. Anything long-running, anything a teammate needs to trigger, anything you want to fire from a phone — all of it stops when you close the lid. A hosted client keeps the connection and the workspace, so a job you start in Discord on the train is still running when you get to a desk, and the files it produced are waiting in the same place.
Register a server with /mcp add, list what is connected with /mcp list, and turn one off with /mcp remove. The same controls are on the dashboard for anyone who would rather click than type.
The client is not tied to one provider. Point it at any model, on your own provider key if you want one, and swap mid-conversation. MCP is an open standard and this is the matching half of that idea.
The tools land inside a full agent loop: it reads files, runs commands in a per-user sandbox, and posts real files back. An MCP tool is one more thing it can reach for, not the whole capability.
The connection runs the other way too. Point Claude Code, Cursor or VS Code at Ergod and call into your own hosted sandbox as a tool: mint a token on the dashboard, set it as a bearer header, and your local agent gets a remote machine that keeps its files between calls. Most MCP servers wrap one API. This one hands you a working environment — and the files it produces are the same ones waiting in Discord and the browser. The server story has its own page: the connect command, all five tools, and the constraints, spelled out.
MCP registrations can be shared into a channel session, so a whole group works against the same connected tools rather than each person wiring up their own. The full technical breakdown covers the sandbox, sub-agents and GitHub linking alongside it.
Free tier, no card, and no local runtime to keep alive.