The MCP control plane
PlatformSmith exposes an MCP server — a Model Context Protocol endpoint that an AI assistant can connect to in order to operate the platform for you: list your workspaces, launch a runtime, follow a session, read a result.
It is a thin proxy in front of the same API you would call yourself. That is the important property:
Everything an assistant does through MCP runs as the user who owns the token, scoped to that user's organization, through the same permission checks. It cannot see or do anything you could not do yourself. → Security and tenancy
Connecting
Point your MCP client at your deployment's MCP endpoint and authenticate with a personal access token. The exact client configuration depends on the assistant; what it needs from you is the endpoint URL and the token.
What it can do
The tools follow the object model, and the scope walk is the same:
workspace → project → environment → { agent definitions, workflow definitions }
Broadly: orient in a tenant (list workspaces, projects, environments); launch runtimes and sessions; follow a session and read its events and artifacts; author agent definitions and playbooks; and diagnose failures.
Configuration resolves inner-over-outer
A value set on a project overrides one set on the workspace, which overrides the organization default. Before assuming what will run, read the effective (merged) value with the preview tools rather than the value at any one scope. → Configuration precedence
Secrets are metadata-only
MCP will list secret schemas, names, references and status, and will wire an agent definition to a secret reference. It never returns a decrypted secret value — a hidden or forbidden secret is omitted or refused outright.
If an assistant tells you it has read a secret's value from PlatformSmith, it has not.
Before you start an agent definition, check it has a credential
Agent definitions carry a flag saying whether their coding agent actually has a credential available to you, in that workspace. If it says no, starting it will be refused.
If the flag is absent, availability is unknown, not available — it is computed per caller and per workspace, so a call that lacks a workspace context cannot answer honestly and deliberately does not guess. → Why a definition can be unlaunchable for you