Connecting agents

Estratos is a standard remote MCP server. Any client that speaks the Model Context Protocol can read and write your company's memory — no bespoke integration. This section walks through connecting each agent your team uses, and how to troubleshoot connections.

How connecting works

Claude installs the Estratos Memory plugin from its marketplace. Codex and other MCP clients point at the memory endpoint directly. Copilot Cowork is real-tenant validated for personal sideload, sign-in, and authenticated memory retrieval; organization-catalog rollout remains owner/admin-gated, and marketplace submission is pending. Once an available client connects for the first time, it runs a one-time OAuth handshake in your browser:

  1. The agent discovers the Estratos authorization server from the endpoint automatically.
  2. Your browser opens the Estratos consent screen.
  3. You pick which workspace to connect (if you belong to more than one) and review the scopes the client is requesting.
  4. You click Allow, and the agent receives a token scoped to that one workspace.

No client ID, secret, or API key goes into any config. The token identifies you and your workspace on every call — you never pass a workspace or account as a parameter.

What installs

The Estratos Memory plugin bundles the MCP memory server plus skills — retrieval, recording, hierarchy, and naming. Clients that connect over MCP directly — Codex and generic MCP clients — get the tools only and rely on the server's built-in tool descriptions.

Everything on this page mirrors the Connect › Configure agents screen inside the app, where each card shows the same steps with a copyable setup command or URL and a live "Test your connection" status.

Pick your agent

  • Claude Desktop — install from the plugin marketplace
  • Codex — one-command setup for the CLI, IDE extension, and app
  • Copilot Cowork — real-tenant validated Microsoft 365 package; organization-catalog rollout and marketplace submission pending
  • Generic MCP client — any MCP-capable agent, manual setup
  • Troubleshooting — reconnecting, token expiry, and testing a connection

Requirements

Connecting an agent requires the API access entitlement on your workspace. Any active member can approve the OAuth consent for their own agent — connecting is per user, and the token binds to a workspace that member belongs to. What a member's agent can do is always bounded by that member's role at the time of each call — see roles and permissions.

Prefer the raw endpoint or a headless setup?

The plugin is the easy path, but everything is a standard MCP + OAuth surface underneath. If you're building your own client, wiring CI, or want the protocol-level details — endpoints, scopes, discovery documents, dynamic registration — see the MCP server reference.