Skip to content
Learning Outcomes
  • Know which AI assistants can talk to Foundry Connect, and how each one connects.
  • Understand why some assistants install a plugin while others need the connection added by hand.
  • Find the onboarding page for the surface you use.

Using Foundry Connect with Copilot, Codex and ChatGPT

This section walks you through talking to Foundry Connect from GitHub Copilot, Codex and ChatGPT: asking about your runs, pipelines and reports in plain language, and launching and monitoring runs. Under the hood it is the same MCP server that backs Claude and every other client; these pages are the per-assistant on-ramp.

Using Claude? That has its own section: Using Foundry Connect with Claude.

The one thing that differs between assistants

Every assistant needs to know your instance address, and this is the only real difference between them. Foundry Connect is not one shared service: each organization runs its own instance, so a published plugin cannot hard code where to connect.

Assistants solve that in one of two ways, and it decides how much setup you do:

  • The assistant fills in the address for you. You install the plugin and set one environment variable. Copilot CLI works this way, as does Claude Code.
  • The assistant cannot fill it in. The plugin still delivers the skill and commands, but you add the connection yourself with one command. Codex works this way, not by oversight but because it stores a bundled server's address exactly as written and never expands a variable inside it.

If a page tells you to add the connection by hand, that is why.

Find your surface

Surface Onboarding path Set it up
GitHub Copilot CLI Plugin, address from instance_host GitHub Copilot CLI
Copilot in VS Code MCP config, address written out Copilot in VS Code
Codex CLI Plugin for skills, connection added by hand Codex CLI
Codex IDE extension No plugin: connection and skills installed separately Codex IDE extension
ChatGPT web Workspace connector, administrator only ChatGPT web
Claude Code, claude.ai, Claude Desktop Plugin or MCP connector Using Foundry Connect with Claude

In this section

  • Getting Started — what you need, the two ways to authenticate, and your first conversational action.
  • GitHub Copilot CLI — install the plugin from the marketplace and point it at your instance.
  • Copilot in VS Code — the extension reads a different file, so it needs its own configuration.
  • Codex CLI — install the skills, then add the connection with one command.
  • Codex IDE extension — no plugin, so the connection and the skills install separately.
  • ChatGPT web — why there is no self-serve path, and what an administrator can do.

Signing in

Two ways to authenticate, the same as on Claude:

  • OAuth, preferred. The assistant is sent to Foundry Connect, you approve access once, and no token is ever stored in a config file.
  • A Personal Access Token, which works everywhere. Generate an MCP token (prefix via_mcp_) under Personal Access Tokens and supply it as a header.

OAuth on Codex needs Foundry Connect 2.22.0 or later

Codex looks for its sign-in details at a different address than other assistants use, and instances older than 2.22.0 do not answer there. On an older instance, codex mcp login cannot complete and you should use a Personal Access Token instead. Copilot and Claude are unaffected, because they follow the address the server hands them rather than guessing one.

See also