Yellow is a place external agents can read and write. Claude Desktop, Claude Code or any client that speaks the Model Context Protocol (MCP) can list your models, search instances, create and update records, run automations and raise proposals, with the same rules a member gets: the agent only sees what its identity allows, and every write becomes a version with its author on it.

This article explains what an agent is in Yellow, how you give it access, and what you see afterwards.

The agent is an application

An agent does not get an account of its own. It acts as an application: the identity an organization creates inside one environment, with roles that say what it may do. You will find applications under the organization, in the «Aplicacions» section.

What the application has What it means for the agent
Environment The agent reads and writes that environment only.
Roles The tools the agent sees are a projection of the roles' capabilities. A role that only allows reading instances gives an agent that can list and read, and nothing that writes.
Read-only With IsReadonly set, every writing tool disappears, whatever the roles say.
Ownership Creating or changing models needs the organization that owns the environment. An application of another organization never gets those tools.

The list of tools an agent gets is what it can actually use: a tool the roles do not allow is not offered, and a call to it is refused before it runs.

Keys

The agent authenticates with a key of the application. Keys live on the application card, in the «Keys» tab.

Field What it does
Label A name to tell keys apart, such as the client that uses it.
Valid for How long the key authenticates: 30, 90, 180 or 365 days. An expired key is refused like a revoked one.
Secret Shown once, when the key is created. Only its hash is stored; if you lose it, create a new key.
Last used When the key last signed in, updated every quarter of an hour at most.

Seven days before a key expires, the owners of the organization receive a notification. Use Rotate on the key: it creates a replacement with the same label, shows the new secret once, and offers to revoke the old key in the same step.

Connecting a client

The application card has a «Connect your agent» section with the endpoint address and a definition to paste for each client. A key's secret only shows once, so the dialog that shows it can copy a definition with the key already in place.

Client How it connects
Claude Desktop Add the entry to claude_desktop_config.json under mcpServers. Claude Desktop only launches local programs, so a small bridge called mcp-remote forwards to the endpoint; it needs Node installed.
Claude Code Run the claude mcp add line once in a terminal. Claude Code speaks to the endpoint directly.
Any MCP client The endpoint address over Streamable HTTP, with the key as a bearer header.
Claude.ai and ChatGPT connectors Hosted connectors sign in with OAuth rather than a key. Paste the endpoint address, sign in with your own Yellow account and choose the environment. The agent then acts as you, with your roles, and every version it writes carries your name.

The endpoint keeps no session, so restarting the backend is transparent to the agent. Each key is limited to 120 requests a minute.

What you see afterwards

Everything the agent does is attributed to the application.

  • Activity: the application card shows the agent's calls per day for a month, its last sign-in, the versions it has written and the proposals it has raised that still await a decision.
  • Usage: the «Ús» panel lists the application under the «Agent (MCP)» source, next to chat and automations, and the organization can set a monthly cap on its tool calls from the application card. In warning mode the cap is logged; in hard mode the agent gets a readable refusal and nothing runs.
  • Proposals: when a model's write policy asks for AI changes to be reviewed, the agent's writes become proposals in the «Propostes» inbox instead of versions.
  • History: every version an agent writes is marked as AI-written, with the application as its author, and can be reverted like any other.

Agents signed in as you

An agent connected through a hosted connector holds a grant in your name, for one environment. Your user settings list them under My agents: the client, the environment, when you granted it, when it last renewed its access and when the grant lapses if unused. Revoke ends the agent's access at its next renewal; it can be granted again by signing in from the client.

Where to find it

Open the organization, go to «Aplicacions», and open the application. The «General» tab holds the roles, the activity and the quota; the «Keys» tab holds the keys and the connector definitions.