Most platform servers let an agent invoke what a human already built. Here the lifecycle operations are exposed too, so a Workflow or Guardrail can be created and revised from inside the session that will use it — which is the difference between an agent that runs your automation and one that extends it. The reporting tools point the same way: usage and effectiveness are queryable, so what gets built next can be informed by what is actually being used.
The hosted MCP server for Oi, reached by URL rather than installed. It exposes Oi's building blocks — Contexts, Workflows, Skills, Guardrails, Connections and durable memory — to an MCP client, covering both using them and authoring them.
- Identity and routing for the calling client
- Contexts, Workflows and available Skills discovered and executed
- Authoring and lifecycle operations over Contexts, Workflows, Skills and Guardrails, so they can be created and changed from the same session that runs them
- Stable Connection discovery and execution — `oi.connections.list`, `oi.connections.get`, `oi.connections.use`
- Durable memory updates at two scopes: Org Brain and User Brain
- Usage reporting for Contexts and Workflows, plus cross-surface effectiveness reporting
An Oi account and a bearer token. Point your client at `https://api.oioioi.ai/mcp` over streamable HTTP with an `Authorization` header — either an OAuth access token obtained through the Oi app, or an exported organization API key. Nothing to install locally. The repository itself holds the registry metadata, published as `oi-mcp` (2.0.0).