Skip to content

Connect an Agent to a Workspace

Every Workspace is exposed as a standard MCP server, so any tool that speaks the protocol can read and write it. You configure the connection once, and from then on the Agent works on its own.

What the connection is made of

Two things, both on the Connection page of the Workspace inside Meridiaan:

  • The address of that Workspace, in the form https://mcp.meridiaan.io/mcp/<workspace id>. Each Workspace has its own, so a client that refuses two servers at the same address can still hold several of them.
  • A token for that Workspace, sent as a Bearer token.

One address and one token, and the client is in:

How a client reaches a Workspace A client calls the address of one Workspace, carrying the token of that Workspace, and from there reaches its Collections and Documents. Your client /mcp/<workspace id> with the Workspace token Workspace Collection

The list of tools travels once, when the connection opens. That is why a Collection created later does not show up in a session that is already running.

Three ways to configure it, from the shortest to the most generic

They are the same thing at three levels of effort, so whoever does not find their own tool goes down one step instead of being stuck.

  1. One click. A button opens the client and creates the entry already filled in. Today it exists for Cursor, VS Code and LM Studio. Two things are worth knowing before you press it: the token travels inside the address of the button, so it ends up in the history of that client, and the button always installs the entry globally, because none of those three accepts a destination.
  2. A ready made block. The configuration already written in the shape your tool expects, and the place where it goes, which is the part that changes most from tool to tool. There are entries for editors, for command line tools, for desktop applications, for those who write the Agent themselves in Python, LangChain or LangGraph, and a generic one.
  3. The minimal formula. The address and the token, named as such, for whoever knows their own client better than we do.

Tools that have been tried

GitHub Copilot, Claude Code and OpenAI Codex have been connected end to end. Anything else that speaks MCP works as well: those three are the ones that have been verified rather than derived from someone else's documentation.

Claude Desktop connects too, from Settings and then Connectors, with a custom connector whose authentication is set to none and an authorization header added by hand. That feature is in beta on Anthropic's side and is available on their paid plans.

Two things worth knowing up front

  • A Collection created while a session is open does not show up in that session. The server does not push changes of the tool list, so the Agent keeps the list it received when it connected. Reconnecting is enough.
  • A token has no expiry today. It stays valid until it is rotated by hand, which is worth remembering when someone leaves the project.

Under the hood

The transport is Streamable HTTP: a single endpoint that answers a POST, with a plain response for almost every call and a stream only where it is needed. Nothing stays open between one call and the next.

On the protocol itself the server is dual era: it negotiates the revision while the connection opens, and answers both clients that stopped at the revisions up to 2025-11-25 and clients on the current 2026-07-28. A client of either era works without you migrating anything.

What is not there yet

Access through OAuth. Today a connection is authenticated with the token of the Workspace, and that is the only way it works.