Connect an MCP server

Meandr verifies it can reach the server and sign in before the server goes live in your catalog. Nothing broken gets exposed to an agent.

The check runs the MCP handshake, notes the protocol version the server speaks and lists everything it offers: tools, resource templates and prompts. If the check fails, the server's page says why, such as an address that doesn't resolve, a certificate that can't be verified or a credential the server turned down. Fix it and retry.

Two gates decide what an agent can call, and this is the first. The server's catalog opens as soon as it's added: tools, resource templates and prompts, each with a switch. What you switch on is what the team allows at all. Off means no workspace can even import it.

Each tool shows its name on the server and the read-only tag it declares about itself. The tag is the server's own claim, so treat it as a hint, not a guarantee. When the server adds tools later, they show up here switched off. Nothing new reaches an agent until you switch it on here and a workspace picks it.

Open the workspace that should use the server and add the server to it. Anyone with write access to the workspace can do this. The credentials stay with the server, so every workspace that adds it reaches the server the same way. No workspace yet? Create one under Workspaces first.

Then decide whether its rate limits are shared with other workspaces or kept separate. Shared counts calls from every workspace together, which fits most servers, since they limit by the key Meandr signs in with. Isolated counts this workspace on its own, for when each workspace has its own budget at the server.

The limits themselves come from the server's settings, or from the workspace's limit per server when the server has none. A team admin can also set Shared or Isolated on the server itself, and then it holds for every workspace.

This is the second gate. Of everything the team allowed, pick the tools this workspace can call. Nothing reaches an agent unless it cleared both.

Names stay as they are upstream. Rename a tool only when two servers collide, as when a workspace uses both GitHub and your own GitHub Enterprise server. You can also give a tool a description of your own. It is what agents read when they decide which tool to use.

Importing none is fine. The workspace's Tools page can do it later, and resource templates and prompts are picked the same way.

That's it. Agents connected to the workspace now see these tools, and every call they make goes through the workspace's policies.

Servers that sign in with OAuth

Pick OAuth · authorization code and sign in to the provider in your browser. Meandr registers itself with the provider when the provider allows it, keeps the tokens encrypted and renews them as they expire. If a renewal ever fails, calls to the server stop and the console asks you to reconnect.

For machine-to-machine access, pick OAuth · client credentials and give Meandr the token URL, a client ID and its secret. Meandr fetches a new token whenever the last one runs out.

One sign-in serves the whole team. Calls from every workspace go out as the account you signed in with, so sign in with an account whose access you mean to share.