Skip to content

MCP security for maintainers

Studio’s MCP server is the one surface Studio deliberately exposes to software it does not control. An AI agent behind it is driven by a language model, and a model can be talked into asking for things nobody intended. The design assumes that.

The promise is narrow and specific: a call made with your organization’s credential can only reach your organization’s data. Everything on this page is either how that is kept, or what you must do yourself because Studio cannot.

Each organization has one MCP credential, made of two parts:

Part What it is Secret?
Client ID Your organization’s name No. It names something your organization already knows
Client secret A generated secret Yes. Whoever holds it acts as your organization

An agent exchanges the pair for a bearer token, and the token is what it sends with each call. The token names the organization it was issued for, and Studio signs and encrypts it, so an agent cannot edit it to name another one. Studio also discards any identity information a client sends along with a request, so a header an agent sets can never pass for one Studio produced.

The organization is not a parameter of any tool. A model that is persuaded to ask for another organization’s model has nothing to ask with.

  • Studio stores the secret encrypted with your organization’s own key, so that members can read it back under Settings > MCP.
  • Studio answers the same way for an unknown client ID, a wrong secret and a revoked credential. Nobody can use the token endpoint to find out which organizations exist or whether one has a credential.

When you ask one of Studio’s agents to do something on a board, it does not reach into Studio. It exchanges your organization’s credential for a token and connects to the public MCP address, exactly as an outside agent does. That is deliberate. Studio’s agents are confined by the same rules as yours, with no second, more privileged path for a model to be talked onto.

Studio ships no AI model of its own. Until an organization has configured a provider and an agent that can reach it, the AI features are off and say why. A reply is recorded as having acted only when a tool was actually called, not when the model says it did something.

Task Where Who
Issue a credential Settings > MCP > Issue credential A maintainer
Read the client ID and secret Settings > MCP, then Show Anyone who can open the page
Rotate the secret Settings > MCP > Rotate secret A maintainer

Every organization is given a credential automatically when its encryption key is created. Issue credential is the way back when that did not happen.

Rotation takes effect immediately. The old secret stops working the moment the new one exists, and every agent that used it has to be reconfigured. Rotate when:

  • the secret was pasted somewhere it should not be, such as a chat, a ticket or a repository,
  • someone who knew it has left the team,
  • you do not know who holds it any more.

A token that was already issued can keep working until it expires, which is an hour after it was issued.

The guarantee is narrower than “safe”. Know the limits:

  • An agent acting inside your organization is fully authorized there. A model talked into removing a module your organization owns is making a legitimate call. Two things stand in the way: the tool descriptions tell the agent to remove only when a person asked, and Studio refuses to remove anything unless every slice in it is still Not started. The tenant boundary does not help here.
  • Anyone who holds the secret is your organization. There is no second factor. Rotation is the remedy, and it is immediate.
  • The secret is shown to people in your organization on purpose. Anyone who can open Settings > MCP can read it. Encryption protects it at rest, not from your own members.
  • MCP access is not read-only. There is no read-only credential today. Connect an agent only when you are comfortable with it changing your models.
  • Keep the secret in a secrets manager or your operating system’s credential store, never in a repository or a prompt.
  • Give each agent its own short-lived token, and let the refresh token renew it, instead of sharing one long-lived token.
  • Rotate on a schedule that matches how often people come and go.
  • Review what agents changed in the board history. Their changes are recorded like anyone else’s.