Connecting an agent over MCP

The MCP screen connects your coding agent, such as Claude Code or Codex, to the platform. Once it is connected, the agent can read your projects and, as more tools arrive, propose changes for you. Everything an agent writes lands in a draft. Nothing it does reaches your live traffic until you publish it yourself.

What it is

MCP is the way your own agents talk to the platform. You create a token on the MCP screen, paste one line into your agent’s config, and the agent can then use the platform’s tools the same way it uses any other.

• A token is how your agent signs in. It belongs to you, carries your access, and can do nothing you could not do yourself in the dashboard. You can hold several at once, one per machine or agent, and name them so you can tell them apart later.
• The tools are what the agent can do once connected. Today it can list your projects, read your catalog, and propose catalog items into a draft (see Defining a catalog item). More tools arrive as the platform grows, and this screen always shows the ones your agent can currently reach. See All MCP tools below for the full set.
• The draft boundary is the rule that never changes. An agent can read, and it can propose changes into a draft, but it can never publish, name a version, change a goal’s value, change rollout settings, manage members, or touch credentials. Those stay with a person in the dashboard.

The draft boundary is why connecting an agent is safe even though the agent reads code and text you did not write. The worst an agent can do is propose a draft you decline.

How to use it

To connect an agent:

1. On the MCP screen, select Create token and give it a name, such as the machine or agent it is for.
2. Copy the token. It is shown once. If you lose it, create another.
3. Copy the config snippet the screen shows and add it to your agent. In Claude Code this is the claude mcp add command on the screen; in other agents it is the server URL and the token in an Authorization header.
4. Start your agent and confirm it lists the platform’s tools. Ask it to list your projects to check the connection end to end.

To rotate a token, select Rotate next to it. The old token stops working at once and you get a new one to paste in. To take an agent’s access away, select Revoke.

Failure modes

• The agent cannot connect. No token was sent, or the one it sent is wrong, rotated, or revoked. The agent’s error tells those two cases apart, so you know whether the config is missing a token or holding a stale one. Create a fresh token and paste it in again.
• A tool the agent expects is missing. The tool list on this screen is the full set your agent can reach. If something is not there, it is not available to agents yet.
• The agent reports it cannot publish, name a version, or change a setting. That is the draft boundary working as intended, not a broken connection. Those actions are yours to take in the dashboard. Publish the agent’s draft when you are happy with it, and its changes go live then.

All MCP tools

Every tool an agent can reach today, grouped by the area it works in. The MCP screen shows the same set for your own token. Reads return data; writes land in a draft, and none of these publish, name a version, or touch credentials.

Tool	What it does
Server
server_info	Report the server name, the tool-surface version, and the member the token acts as.
Models
list_models	List the organization's models (id, name, provider, model id). Keys are never returned, and configuring models is dashboard-only.
Catalog
list_catalog	List the organization's catalog items, each with its id, name, component identifier, and latest published version.
get_catalog_item	Read one catalog item by its component identifier, with every version and its properties.
upsert_catalog_item	Define a catalog item into its draft from a full definition, creating it if it does not exist.
Tools
list_tools	List the organization's tools, each with its id, tool name (the function name the model calls), display name, owning project, scope, whether it runs at session start, and when it last changed.
register_tool	Register a tool into its draft from a full definition, creating it if the tool name is new.
System prompt
list_system_prompt_versions	List a project's system prompt versions, each with its number, status, and text.
get_system_prompt_version	Get one system prompt version, including its text.
create_system_prompt_version	Start a new draft system prompt version, seeded from a source version or empty.
edit_system_prompt_draft	Replace the text of the draft system prompt version.
Projects
list_projects	List the organization's projects (id and name).
create_project	Create a project with a name unique to the organization.
update_project	Rename or re-describe a project. Refused once the project has a published version.
Project versions
list_project_versions	List a project's versions, each with its status and the model, system prompt version, and catalog item versions it pins.
get_project_version	Get one project version, including everything it pins.
create_project_version	Start a new draft project version pinning a model and system prompt version, optionally with catalog item versions.
update_project_version	Edit a draft project version's model, system prompt version, description, or pinned catalog items.