Paste one URL and sign in

https://mcp.megamcp.ai/c/{connectionId}

Example:
https://mcp.megamcp.ai/c/conn_a1b2c3

Find yours in the dashboard: open your connector and look for the Connector URL row on the Overview tab, with a copy button beside it. In Team mode that row is the admin connection, and each teammate has a personal URL on the Members tab.

On Free, Pro, or Max

1. Open Claude in a browser and go to Settings, then Connectors.
2. Click Add custom connector.
3. Paste your connector URL into the server URL field.
4. Leave Advanced settings alone. The OAuth Client ID and Secret there are optional and megamcp does not need them. Leave the request-header fields empty too.
5. Click Add, then Connect. A megamcp page opens asking you to approve access. Approve it and Claude returns to the connector list.
6. In any chat, use the + button, choose Connectors, and switch your connector on for that conversation.

On Team or Enterprise

Only an Owner can add a custom connector, and adding it is a separate act from connecting to it. Every member signs in as themselves afterwards, which is what keeps a Team-mode connector honest about who did what.

1. An Owner opens Admin settings, then Connectors, and clicks Add custom connector.
2. The Owner pastes the connector URL and clicks Add. This does not sign anyone in. It only makes the connector available to the workspace.
3. Each member then opens Settings, then Connectors, finds the entry labelled Custom, and clicks Connect to approve access for themselves.
4. Members enable it per conversation from the + button, the same as everyone else.

1. Run the add command in a terminal with your own URL in place of the example below. Nothing is validated at this point. The command only writes configuration, so a bad URL stays quiet until step three.
2. Decide where it lives. With no scope flag the connector is added for the current project only. Add --scope user to make it available in every project on your machine, or --scope project to share it with your team through a checked-in .mcp.json.
3. Start Claude Code and run /mcp. Pick megamcp from the list to sign in, and your browser opens the megamcp approval page. You can also run claude mcp login megamcp straight from the shell.
4. Confirm it worked with claude mcp list. A healthy connector reads Connected. Needs authentication means step three is unfinished, and Failed to connect almost always means the URL is wrong.

# Add it (this project only)
claude mcp add --transport http megamcp https://mcp.megamcp.ai/c/conn_a1b2c3

# Or make it available in every project on this machine
claude mcp add --transport http megamcp --scope user https://mcp.megamcp.ai/c/conn_a1b2c3

# Then, inside Claude Code:
/mcp

1. Turn on developer mode first, or no custom-server option appears at all. OpenAI documents this under Settings, then Security and login. It is a web-only feature on Pro, Plus, Business, Enterprise, and Education accounts. ChatGPT Free cannot add one.
2. On Business or Enterprise, an admin may have to allow connected data for the workspace before the toggle does anything. If developer mode will not stay on, that is the reason.
3. Create a developer-mode app for a remote MCP server and paste your connector URL as the server address.
4. Choose OAuth as the authentication method. Do not choose no authentication. ChatGPT rejects a server quietly when the chosen auth mode does not match what the server actually does, and megamcp always requires authentication.
5. Approve access on the megamcp page when ChatGPT sends you there. The connector then appears under the composer's developer-mode tool inside a conversation.

Create one from the connector's Settings tab. It is shown exactly once, because only an encrypted form is kept, so put it in your secret manager the moment you see it. It travels as an Authorization: Bearer header on the same connector URL, and it can be rotated or removed later without rebuilding anything.

curl -X POST https://mcp.megamcp.ai/c/conn_a1b2c3 \
  -H "Authorization: Bearer $MEGAMCP_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

# Claude Code, with a static token instead of signing in
claude mcp add --transport http megamcp https://mcp.megamcp.ai/c/conn_a1b2c3 \
  --header "Authorization: Bearer YOUR_TOKEN"

The official MCP client library still defaults to the older protocol revision for backward compatibility. Your connector serves that revision happily, so setup, tool listing, reads, and dry runs all look fine. Writes that pause for human approval are the exception: the approval round trip does not exist in the older revision, so those calls are refused outright with a message about upgrading the client. It reads like a broken connector and it is a one-line client setting. Because it is a client-side default, it can only be fixed in your code.

const client = new Client(
  { name: "my-agent", version: "1.0.0" },
  {
    capabilities: { elicitation: {} },
    // Without this, the client speaks the older revision and every
    // approval-mode write is refused. Use "auto" to probe and fall back,
    // or { pin: "2026-07-28" } to fail loudly instead.
    versionNegotiation: { mode: "auto" },
  },
);

That connection is not available on this account

You are signed in to megamcp as the wrong person, or in a browser where you are not signed in at all. The approval page needs the megamcp account that owns the connector, in the same browser. Sign in, then start the connection again from the client rather than reloading the approval page.

Sign-in never opens, or the client says it needs authentication

Nearly always the URL. Remove the connector from the client, copy the URL again from the dashboard with the copy button, and add it fresh. Check for a trailing slash and for a stray space at either end.

It connected, but no tools appear

Connecting and enabling are two different things. In Claude, switch the connector on for the conversation from the + button. In ChatGPT, enable it in the developer-mode tool. In Claude Code, run /mcp and check the tool count next to the server.

Reads work and every write fails

If the failure asks a human to approve something, that is the connector working as designed: writes pause on a preview and wait for a person. If the error says the approval round trip needs a newer protocol revision, your client is speaking an older one. See the section on writing your own client.

Sign-in worked but every call fails against your software

There are two sign-ins and only one is done. Open the connector's Settings tab in the dashboard and connect your account for the upstream software. Until that is finished the connector authenticates perfectly and reaches nothing.

A browser-based tool gets a 403

The endpoint does not serve browser origins, deliberately. Anything calling it from inside a web page is refused before authentication even runs, which is why an in-browser inspector fails while real clients work. Test with the AI client itself, with Claude Code, or with curl.

Everything is refused at once

Check whether the connector, or that one member's connection, is paused or revoked in the dashboard. Both take effect on the very next request, with nothing cached that could serve past them. A lapsed subscription has the same effect.