STACK with ChatGPT Workspace Agents

Why this matters

ChatGPT Workspace Agents lets a workspace admin build agents that take action on connected services, schedule tasks, run on a credit budget, and stay inside the controls the admin sets (RBAC on who can publish, which connectors are enabled, when an action needs approval). For the first 90% of an agent's life (chatting, reasoning, hitting OpenAI's curated connectors), that's a tight setup.

STACK shows up the moment an agent needs to do something OpenAI's connector set doesn't natively cover: post to a Slack channel that isn't on the curated list, fire a webhook into your billing pipeline, transfer money, sign a contract. Whatever the agent reaches outside OpenAI's walls runs through one chain you control.

What STACK opens up extends beyond Workspace itself. Your other agents (on Claude, Cursor, custom runtimes) sit on the same identity, the same audit chain, the same kill switch. And every agent in that fleet, whichever runtime it lives on, can hand work off to the next one without breaking the chain.

What it unlocks

Three things STACK adds to a Workspace Agents deployment that the Workspace toolkit alone doesn't cover today:

• Reach beyond the curated connector catalog. Workspace lets the admin pick from OpenAI's catalog and add custom MCP servers. STACK is the custom MCP that opens up everything else: any HTTPS service, any internal tool, any service you wouldn't or couldn't put on OpenAI's list. Same admin RBAC on top, plus STACK's own audit and revocation underneath.
• Credentials in your vault, not OpenAI's. Workspace's connector flow stores connected-account credentials in OpenAI's vault. For regulated workloads where the credential cannot leave your control (financial, healthcare, internal enterprise), STACK keeps it KMS-encrypted in your own vault with your own RBAC. The agent never sees the secret. OpenAI never sees the secret.
• Cross-runtime, cross-team hand-offs. Workspace Agents are workspace-scoped — an agent in one workspace can't naturally pass work to an agent in another, or to an agent running on Claude or in your own runtime. STACK's drop-offs let an agent hand a schema-validated payload to the next agent (any team, any LLM, or a published skill from the marketplace) with the audit lineage preserved end to end.

Wiring it up

Add STACK to the Workspace Agents builder as a custom MCP connector. STACK speaks OAuth 2.1 with PKCE and Dynamic Client Registration, which is what the builder requires — no API key copy-paste; users approve scopes in their browser. Step-by-step setup, the MCP URL, available scopes, and what happens during the OAuth handshake:

→ /docs/integrations/openai-workspace-agents

Last reviewed 2026-05-08. ChatGPT Workspace Agents available on Business, Enterprise, Edu, and Teachers plans (research preview). Free until 2026-05-06; credit-based pricing thereafter.