Client recipes

ChatGPT with Stable Baseline.

OpenAI's ChatGPT exposes remote MCP servers through Developer Mode inside Apps & Connectors. Flip the toggle, paste the Stable Baseline URL, and chat against your entire workspace.

Heads-up — this is a beta flow

Custom MCP servers live under the Apps & Connectors surface in ChatGPT, behind the Developer mode toggle. As of November 2025 it's available on all plans — Free, Plus, Pro, Business, Enterprise, and Edu — but remains a beta feature. On Business, Enterprise, and Edu plans a workspace admin has to turn it on from Workspace Settings → Permissions & Roles → Connected Data → Developer mode / Create custom MCP connectors before individual users can flip it on.

Before you begin

You'll need two things:

A Stable Baseline API key — see the MCP setup guide. Keys are prefixed sta_… and shown only once.
A ChatGPT account. Developer mode is available on every plan, but if you're on Business / Enterprise / Edu your admin needs to allow it at the workspace level first.

Step 1 · Enable Developer mode

Open chatgpt.com. Click your avatar (top-right) and pick Settings.
Go to Apps & Connectors.
Scroll to the bottom of the page and open Advanced settings.
Turn on the Developer mode toggle. ChatGPT shows a beta-feature warning — accept to continue. A Create button will now appear on the Apps & Connectors page.

Developer mode relaxes some guardrails

OpenAI explicitly flags that custom MCP servers run without some of ChatGPT's standard tool-call safety filters — only connect servers you trust. Stable Baseline's MCP server never reads or writes outside the workspace the API key is scoped to.

Step 2 · Create the Stable Baseline connector

Back on Settings → Apps & Connectors, click the new Create button.
Fill in the Create dialog:

ChatGPT — Create connectortext

Connector name:  Stable Baseline
Description:     Stable Baseline workspace — docs, diagrams,
                 plans, improvements, compliance. Use for
                 workspace search + writes.
Connector URL:   https://api.stablebaseline.io/functions/v1/cloud-serve/mcp
Authentication:  OAuth

Save. ChatGPT opens the OAuth consent window, where you sign in to Stable Baseline and authorise the connector. After approval, ChatGPT fetches the tool manifest and the connector shows as Connected.

API key vs OAuth

The Create Connector dialog expects OAuth 2.1 + dynamic client registration (the MCP spec's recommended auth path), which Stable Baseline supports natively. If your workspace policy requires bearer-token auth instead, use the mcp-remote stdio bridge from one of the other clients (Claude Code, Cursor, OpenCode) and point ChatGPT at that machine — but OAuth is the simplest path.

Step 3 · Enable per-chat

Custom connectors aren't on by default in every conversation. When you start a new chat:

Click the + button next to the message composer, then pick More.
Under Developer mode you'll see every custom connector you've created. Tick Stable Baseline.
ChatGPT remembers the choice for that conversation. Every new chat where you want to use Stable Baseline needs the toggle flipped once.

Step 4 · Verify the connection

Type a quick probe into the chat:

ChatGPTprompt

Using the Stable Baseline connector, list my workspaces.

ChatGPT will call listWorkspaces and render the result. Any action that writes or modifies data triggers a confirmation modal before execution — expected behaviour, accept to continue.

Nothing coming back?

OAuth expired — disconnect and re-authorise from Apps & Connectors.
Connector not enabled for this chat — open + → More and tick it on.
Workspace policy — if you're on Business / Enterprise / Edu, confirm your admin has turned on custom MCP connectors at the workspace level.

Step 5 · Try a real task

“Find all open improvements in the Platform project with priority critical and summarise the blockers.”

“Read the PRD for Multi-region GA and draft a launch checklist. Save it as a new document in the Product folder.”

“Create a sequence diagram for our login flow and insert it into the Authentication doc.”

What works, what doesn't

Tool calls — full 60+ MCP tools are exposed, same as any other client.
Write confirmations — ChatGPT shows a modal before every action that writes or modifies data. No way to bypass — it's a deliberate safety guardrail for custom connectors.
Long context — ChatGPT's context window applies. For bulk documentation passes, an IDE-based agent (Claude Code, Cursor) is usually faster and cheaper.
File uploads — attaching a PDF or image in the ChatGPT composer does not flow through the MCP server automatically. To attach a file to a Stable Baseline document, ask the agent to create an upload session via createImageUploadSession.

Security notes

The OAuth token is stored by OpenAI in your ChatGPT account and scoped to the workspace your Stable Baseline user has access to.
Use a project-scoped connection whenever possible so the blast radius is limited to one project's documents.
Developer mode is explicitly off by default. Turning it off in Settings also disables every custom connector you've created.

Claude Code recipe

Anthropic's agentic CLI.

Gemini recipe

Google's chat + CLI over MCP.

Permissions

Scope per agent, not per user.

Tools reference

Every MCP tool available.