MCP server: connect Claude and ChatGPT

Put the token in the WLAUNCH_MCP_TOKEN environment variable, loaded from a secret manager or your system keychain, then add the server from your terminal:

claude mcp add --transport http wlaunch https://mcp.wlaunch.net/mcp \
  --header "Authorization: Bearer $WLAUNCH_MCP_TOKEN"

Check that it was added with claude mcp list, and remove it with claude mcp remove wlaunch. Do not type the token itself into the command: the shell keeps every command in its history. Claude Code writes the token into its config in plain text, so treat that file as a credential.

With the token in the WLAUNCH_MCP_TOKEN environment variable, add the server:

codex mcp add wlaunch \
  --url https://mcp.wlaunch.net/mcp \
  --bearer-token-env-var WLAUNCH_MCP_TOKEN

Or add it to ~/.codex/config.toml:

[mcp_servers.wlaunch]
url = "https://mcp.wlaunch.net/mcp"
bearer_token_env_var = "WLAUNCH_MCP_TOKEN"

bearer_token_env_var names an environment variable rather than holding the token: Codex reads it at connection time and sends Authorization: Bearer <access_token>. Prefer it to writing the token into the file. It is the one client here that keeps the token out of config and out of logs: load the token into the environment from a secret manager or your system keychain, never as a literal in a shell profile or a repository.

http_headers (a fixed value) and env_http_headers (a value read from an environment variable) also exist, but this server reads only the Authorization header, so bearer_token_env_var is the one to use.

Add the server as a custom connector. The token goes in the Request headers section of the Add custom connector dialog.

Anthropic offers Request headers as a beta to a limited set of organizations. If your dialog has no Request headers section, your account does not have it yet and Claude Desktop cannot send this server your token: connect through Claude Code instead.

On Team and Enterprise plans only an Owner can add a custom connector, under Organization settings → Connectors, and a header stored there is one credential shared by everyone in the organization. That goes against one token per person, so on those plans connect through Claude Code instead.

The order of the steps below matters: set the authentication first, then the header.

1. Open Customize → Connectors and press Add custom connector.
2. Enter Wlaunch as the Name (it is what appears in the connectors list) and https://mcp.wlaunch.net/mcp as the URL, then press Continue.
3. Under Authentication, choose No sign-in. The dialog may detect Sign in now and preselect it: do not take it. The server publishes no OAuth discovery metadata (every /.well-known/oauth-* path returns 404), so a sign-in flow has no authorization server to reach and fails. With No sign-in selected, the OAuth client section no longer applies.
4. Under Request headers, press Add header. The header name is a dropdown: pick authorization and enter Bearer <access_token> as the value. Leave Required checked. Include the Bearer scheme in the value: a bare token without it is rejected. The server reads no other header name, so x-api-key and the rest are ignored.
5. Press Add. The dialog stores header values without showing them again, so keep the token somewhere you can retrieve it.

If authorization is grayed out in the dropdown, you are still on Sign in now. In that mode Claude fills this header itself and does not let you override it. Go back, choose No sign-in and open the dropdown again.

After it connects, the connector page lists Tool permissions: 28 read-only tools, the staff diagnostics ones included. Each can be set to Always allow, Needs approval, Blocked or Custom, one by one or for the whole group. Every tool only reads, but its results include text your clients typed, such as names and comments, and such text can carry instructions aimed at the assistant. Choose Always allow only if this assistant has no tools that can send data elsewhere; otherwise keep Needs approval, at least on wlaunch_api_get.

In the ChatGPT app, custom connectors live under Settings → Connectors, on the plans that offer Developer Mode. Add the server by URL:

https://mcp.wlaunch.net/mcp

ChatGPT's connector dialog is built around OAuth or no authentication, while this server needs a fixed bearer token. If the dialog gives you no field for a custom header, this combination is not supported, and the Responses API below is the route that works. We have not been able to confirm which of the two the current dialog does, so try the app and fall back to the API.

The OpenAI Responses API takes an MCP server as a tool and lets you set headers, which is what this server needs. With the token in the WLAUNCH_MCP_TOKEN environment variable, send:

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d @- <<EOF
{
  "model": "gpt-5",
  "tools": [{
    "type": "mcp",
    "server_label": "wlaunch",
    "server_url": "https://mcp.wlaunch.net/mcp",
    "headers": { "Authorization": "Bearer $WLAUNCH_MCP_TOKEN" }
  }],
  "input": "What companies can this integration see?"
}
EOF

require_approval controls whether the model pauses for your confirmation before each tool call. The example leaves it out, so every call waits for your approval. Setting it to "never" is an option only for an assistant with no tools that can send data elsewhere: tool results include text your clients typed, such as names and comments, which can carry instructions aimed at the model.

Your Wlaunch token goes to OpenAI in that header. That is the same trust decision as pasting it into any hosted assistant.

Use the Streamable HTTP transport, the URL https://mcp.wlaunch.net/mcp and the Authorization header. There is nothing Wlaunch-specific in the protocol, so any client that speaks MCP over HTTP and lets you set a header will work.