Connect to ChatGPT

Core installation and ChatGPT access are separate. With the npm installer, run openmachine once and check openmachine status for ready. For a source installation, complete initial configuration and keep serve running. Either way, verify that your account/workspace permits the intended MCP route. Do not enable machine writes to solve a connection failure.

Select the supported route

For a private custom connection, use OpenAI's separately managed Secure MCP Tunnel. The upstream guide owns current installation, account eligibility, credentials and UI instructions. Core neither creates a tunnel nor enrolls your account.

You need permission to add/use custom MCP servers in the target ChatGPT workspace. Tunnel creation requires the Platform tunnel management permission; operating or selecting one requires its use permission. Associate it with the intended organization and ChatGPT workspace. These are not Core configuration settings.

The tunnel route is for private connections and testing, not a claim that Core is already distributed as a public plugin. Public plugin distribution has separate endpoint and authentication requirements in the upstream guide.

Bind the tunnel to Core

Install and configure tunnel-client according to its official guide, keeping credentials in supported local references. Do not paste secrets into ChatGPT, shell arguments, screenshots or repository files. Inspect the installed client's help because its flags can evolve independently of Core.

Core's existing Unix-socket binding is:

--mcp.server-url=channel=main,url=http://localhost/mcp,unix-socket=/absolute/path/to/state/mcp.sock

Replace the socket path with the same state directory used by serve or the managed service. This is a server-binding argument, not a complete tunnel installation command. Preserve the logical host localhost and Host/Origin protection. Core opens no TCP port; an arbitrary localhost:8000 URL is not its endpoint.

Use the tunnel client's documented readiness checks and keep it running. Do not repoint an existing production tunnel at a disposable candidate just to test it.

Add and approve the connection

Follow the upstream Connect from ChatGPT instructions: add a custom MCP server, select the Tunnel connection and the intended tunnel, then review its authentication and risk prompts. UI availability and wording can vary. Missing access requires the responsible account/workspace administrator, not a Core patch.

After creation, enable the connection in the intended conversation and check that the client actually exposes the expected Core tools. Approving a connection does not mean that every future mutating call is approved.

Prove one real read

For the source-install demo, paste this into the connected conversation:

Use OpenMachine for a read-only connection check. Read its operating Skill and
get_capabilities. Report the actual Core version, then read README.md from
machine local, workspace project. Return its text and complete SHA256.
Do not create tasks, run commands, write files or change any configuration.

For the source-install demo, the returned text must be OpenMachine connection check followed by a newline. Compare its hash locally:

python3 -c 'import hashlib, pathlib, sys; print(hashlib.sha256(pathlib.Path(sys.argv[1]).read_bytes()).hexdigest())' "$DEMO/project/README.md"

For an npm installation, use the same read-only check on machine local, workspace project, but request the existing file chosen during setup (replace README.md in the prompt if needed). Match its actual contents and SHA256 locally, for example with shasum -a 256 /absolute/project/README.md using your real file path. The $DEMO path and marker text above apply only to the source-install demo.

A real tool response with matching bytes/hash establishes this read path. Discovery, a healthy tunnel or a local diagnostic alone does not. Continue with First task; failures belong in Troubleshooting.