Give ChatGPT access to skill discovery through one MCP connection. Describe your task, and your agent can find and load relevant instructions without installing each skill separately.
Before you start
Use ChatGPT on the web with permission to create custom MCP connections. OpenAI documents full support for Business and Enterprise/Edu workspaces; availability and creation permissions depend on your plan and administrator.
You need a 9octopus account to approve access. You can sign in or create one during connection. Use the account for this environment; staging and production accounts are separate.
1. Add 9octopus to ChatGPT
Use this server URL with remote MCP over Streamable HTTP and OAuth:
https://app.9octopus.work/mcp- Open Plugins. Select the plus button, then Add custom MCP server. The current Plugins flow does not require Developer mode. If your workspace still uses legacy Apps, use Settings → Apps → Create instead.
- Configure 9octopus. Name the connection 9octopus, enter the server URL above, and choose OAuth authentication. Review the displayed permissions and warning, then create the plugin. In legacy Apps, scan the tools before selecting Create.
The examples name the connection 9octopus. If you choose another name, use that name in commands and prompts. If you already have a connection, inspect it before adding another.
2. Sign in and approve access
Complete the 9octopus sign-in and consent prompt, then return to ChatGPT to finish discovering the tools.
On the 9octopus consent page, check the client name, sign in, and approve access. Keep the agent open until it finishes connecting. You do not need to paste a bearer token into your configuration.
3. Check the connection and try a task
Open a conversation and enable or select the 9octopus plugin/app for that chat. Check that its tools are available before sending the example task below.
You can also open Connected agents in 9octopus to confirm that the connection is active. Then send this prompt to ChatGPT:
Use the 9octopus MCP connection to find public skills for reviewing a landing page’s accessibility. Show me the relevant descriptions and sources. Load the best fit, confirm its exact version, and explain what access you need before starting the review.
Your agent should discover skills, load the selected release, and read its instructions. It still needs the tools and permissions required to perform the task. If no suitable skill exists, refine the task or browse public skills.
Public task recommendations may send your task description to OpenAI. Keep secrets and private repository contents out of discovery prompts.
Optional: help your agent discover skills as you work
Open Connections and select Agent instructions. Copy the guidance into your client’s persistent instructions, or merge the downloaded AGENTS.md into a supported instruction file. Keep existing rules; supported filenames and settings vary by agent.
Troubleshooting
- If creation is unavailable, ask your workspace administrator about custom MCP access. ChatGPT must be able to reach the public HTTPS endpoint; a local-only URL will not work for this setup.
- Expired consent or HTTP 401: restart authentication from the agent and complete the new browser flow. An old consent URL cannot be reused.
- Pending in Connected agents: the agent has not finished the token exchange. Retry sign-in, then revoke any abandoned pending entry.
- Need to disconnect? Revoke this connection in Connected agents. To reconnect later, clear its saved authentication in the client and sign in again.
Setup reference: ChatGPT MCP documentation. Instructions reviewed October 27, 2026; client menus and plan availability can change.
Using a different agent? Choose another setup guide or follow the general MCP connection instructions.