Skip to main content
TrevoSDK MCP is a hosted MCP server at https://mcp.trevosdk.com/mcp. It lets an AI app do what the CLI does: list what Trevo proposed, approve or dismiss, ask for a fresh batch, follow generation, and start, end and decide experiments. Nothing runs on your machine.

Connect

Add the server to your app. On first use it opens a Trevo sign-in, then a page where you choose:
  • What it can reach. All your organizations and workspaces (the default), one organization, or one workspace. All and one organization include ones you join later, and drop ones you leave, without connecting again.
  • What it may change. Every write is off until you turn it on, and in each organization it can only do what your role there allows. Reading is always included.

Claude Code

Then run /mcp in Claude Code and choose Authenticate.

Cursor

Add it to ~/.cursor/mcp.json, or .cursor/mcp.json in a project:

claude.ai

Settings → Connectors → Add custom connector, with the URL above.

ChatGPT

Add a custom connector with the URL above and OAuth as the authentication. Custom connectors need developer mode, which depends on your ChatGPT plan.

With an API key

For a script, a CI job or a headless agent, send a workspace API key instead of signing in. Make one under Settings → API access:
A key opens its own workspace with its own scopes, as it does for the CLI. Apps that only support OAuth, such as ChatGPT, cannot use one.

Switch workspaces

Tell the app which workspace you mean, for example “use Storefront in Acme”. It calls workspaces_list and passes that workspace’s id to every other tool, so one connection covers all of them. With one workspace in reach, it needs no id. Each call acts in one workspace; asking about all of them at once takes one call per workspace. Working for several clients? Connect once per client, each reaching that client’s organization only. Data from one client then never sits in an agent that can change another’s.

Tools

Every write is marked as one that changes things, and the server asks the app to confirm with you first. generate and proposals_approve answer before the work is done: follow them with runs_show or experiments_status, which wait up to 45 seconds per call.

What it can do

The app acts as you, with the writes you allowed and your role in that workspace’s organization still holds. A write it may not make gets Trevo’s refusal, with the reason. A workspace outside its reach reads as missing. Each workspace’s audit log records when an app that reaches it is connected or disconnected, and each change under your name, marked with the app that made it, for example “via ChatGPT”. What the app reads from Trevo goes to it and to its model provider, and is handled under that app’s terms. Connect only apps your company allows for this data.

Disconnect

  • Account → Connected apps lists the apps you connected, with what each one reaches.
  • Settings → API access → Connected apps lists the apps that can reach that organization. Anyone who can manage API keys sees and disconnects everyone’s. An app connected to all of a member’s organizations is disconnected from this organization only; their other organizations keep it until the member connects it again.
A disconnected app stops at once. Leaving an organization stops every app you connected to it; an app connected to all your organizations keeps working in the others. If you rejoin, they work again until you disconnect them.