MCP server

Connect your AI client to Tivya

Tivya runs a remote MCP server, so an AI client can read and build on your boards directly — no key to generate, nothing to install. Paste one URL into Claude, Claude Code or any client that speaks MCP, sign in to Tivya, approve it once, and the agent works on the same canvas your team does. Everything it does goes through the same operation engine and the same permission checks as a person clicking in the app.

The server URL

This is the only thing you need. Paste it wherever your client asks for a remote MCP server or a connector URL.

https://api.tivya.space/mcp

Want it read-only? Say so when you approve it

There is one URL, on purpose. Read-only is not a different address to remember and mis-paste — it is a choice on the approval screen: pick Viewer and the connection can read your boards and change nothing, ever. Pick Editor and it can build. You can revoke and reconnect at a different role at any time.

Setting it up

Every client asks for the same thing — the URL above — in a slightly different place.

Claude Code

Tested

Run this in any project. Claude Code then asks you to authenticate, opens the approval screen in your browser, and the tools appear in that session.

claude mcp add --transport http tivya https://api.tivya.space/mcp

Claude (web and desktop)

Settings → Connectors → Add custom connector, then paste the URL above. Claude opens the Tivya sign-in and approval window itself; once you approve, the Tivya tools are available in your chats.

Cursor

Add Tivya to your MCP configuration — Settings → MCP → Add new MCP server, or paste this into your mcp.json.

{
  "mcpServers": {
    "tivya": { "url": "https://api.tivya.space/mcp" }
  }
}

VS Code

Run this once and VS Code adds Tivya as an HTTP MCP server for Copilot's agent mode. You can also add it by hand through the MCP: Add Server command.

code --add-mcp '{"name":"tivya","type":"http","url":"https://api.tivya.space/mcp"}'

Any other MCP client

Tivya is a plain OAuth 2.1 resource server over Streamable HTTP, following the current MCP authorization spec: protected-resource metadata, dynamic client registration, authorization code with PKCE, and resource indicators. A client that implements that spec needs nothing from us but the URL — it will discover the rest on its own from the 401 it gets first.

We only mark a client as tested once we have actually driven this server from it. The rest follow each client's own documented way of adding a remote MCP server over HTTP, and the protocol underneath is the same one the tested clients use.

Signing in and approving

Four steps, and you only do them once per client.

1Paste the URLYour client asks Tivya what it needs, is told it needs authorization, and registers itself. Nobody issues it credentials by hand.
2Sign in to TivyaA browser window opens on Tivya's own sign-in — the one you already use, Google included. Your password is never shown to the client.
3Approve itThe approval screen names the client asking, the workspace being connected, and what it may do. You pick the workspace and the role.
4That is itThe client gets a short-lived token bound to Tivya alone, and refreshes it quietly from then on. Nothing is stored in a config file.

A connection can only ever do less than you can

The role you grant is capped at the role you hold — an editor cannot connect an agent as an admin. And the cap is re-checked on every single call, not just at approval: if your role is lowered later, the connection narrows with it the same second. It reaches one workspace and no other, and it can never change your team, its people, its plan or its settings.

Every connection is revocable, and you can see what is connected at any time in Admin → Connections

What it can do

Seventeen tools, and the same ones the local server ships. Everything that changes a board goes through Tivya's operation engine, so it appears in board history and can be undone.

Read my boards

List boards, open one, look at its objects, search it, or ask Tivya's own AI to summarise it.

list_boards
get_board
get_board_objects
find_objects
summarize_board

Create and change things

Stickies, shapes, notes, frames, connectors and groups — created, edited and deleted by id or by exact title.

create_object
update_object
delete_object
create_frame
connect_objects
group_objects

Arrange and lay out

Move something, place it relative to something else, or build and lay out a whole flow or architecture diagram in one call.

move_object
move_relative
apply_layout
create_flow
create_architecture

Undo

The last batch an agent applied comes back off the board as one atomic change.

undo_last

Questions

The ones worth answering before you connect anything.

Which workspace does it use?: The one you picked when you approved it, and only that one. A tool call naming another workspace is refused outright rather than quietly ignored. If you want an agent on a second workspace, connect it again and approve it there — each is its own connection, with its own role, revocable on its own.
What happens if I leave the workspace, or my role changes?: The connection follows you down immediately. Your membership is read from the database on every call, so an agent connected as an editor stops being able to write the moment you are made a viewer, and stops working entirely when you leave. There is no stale token that keeps working.
Is my board content sent anywhere?: Your AI client reads whatever you ask it to read, the same way it reads a file you open — that is the point of connecting it. Tivya itself sends nothing to a third party for these tools, with one exception you ask for by name: summarize_board calls Tivya's own configured AI provider, exactly as the AI panel in the app does.
How do I revoke it?: Admin → Connections, then Revoke. It stops working that second — every token the connection holds dies with it, and the client would have to ask again and be approved again.
My client does not support remote MCP servers. Now what?: Tivya also ships a local stdio server for exactly that case. It runs on your own machine and authenticates with a workspace API key you generate in the admin console. It is the same seventeen tools; only the transport and the sign-in differ.
Do I need a paid plan?: No. The MCP server is open to every workspace, on every plan, while Tivya is in its early days.