Create a key in Dashboard → API Keys, then point your client at the hosted endpoint. Nothing to install, and it stays current without anyone upgrading a package.

https://uni-msg.com/mcp        Authorization: Bearer <key>
https://uni-msg.com/mcp/<key>  for clients that cannot send headers

Claude Code

claude mcp add --transport http uni-msg https://uni-msg.com/mcp \
  --header "Authorization: Bearer wsa_your_key"

Cursor — `~/.cursor/mcp.json`

{ "mcpServers": { "uni-msg": {
  "url": "https://uni-msg.com/mcp",
  "headers": { "Authorization": "Bearer wsa_your_key" }
} } }

Gemini CLI — `~/.gemini/settings.json`

httpUrl selects Streamable HTTP; url would mean SSE.

{ "mcpServers": { "uni-msg": {
  "httpUrl": "https://uni-msg.com/mcp",
  "headers": { "Authorization": "Bearer wsa_your_key" }
} } }

Claude Desktop

Settings → Connectors → Add custom connector, URL https://uni-msg.com/mcp, header Authorization: Bearer wsa_your_key.

ChatGPT

Developer mode (Settings → Apps → Advanced), then add a connector with URL https://uni-msg.com/mcp/wsa_your_key and No authentication. It is the only client that cannot send a header, which is why the key sits in the path — and why nginx disables access logging on /mcp.

Running it locally

Not required — the hosted endpoint above needs no install. Use this if you would rather the API calls left your own machine.

{ "mcpServers": { "uni-msg": {
  "command": "npx",
  "args": ["-y", "uni-msg-mcp-server"],
  "env": { "UNIMSG_API_KEY": "wsa_your_key" }
} } }

Needs Node 20+. ChatGPT cannot use this form — it has no way to run a local process.

Pairing a number

Three steps, and the tool descriptions walk the assistant through them:

create_device with a name and the number in international format
connect_device returns an 8-character pairing code
On the phone holding that number: WhatsApp → Settings → Linked devices → Link a device → Link with phone number instead → type the code

The code expires in about two minutes. If it lapses use request_pairing_code rather than reconnecting — it is cheaper and avoids churning the session. Poll get_device_status until it reports connected.

Notes

Sends are immediate and cannot be recalled. The tool descriptions instruct the assistant to confirm the recipient first, but that is guidance, not a guarantee — treat this as you would a shell with production access.

Only a device whose status is connected can deliver. list_devices first saves a confusing failure later.

Reading message history is exposed: list_conversations finds the peer value, then read_thread returns that thread newest-first with paging.

Errors come back as tool results rather than thrown exceptions, so the model can read the reason and correct itself. Anything key-shaped in an error is redacted before it reaches the model. Diagnostics go to stderr — stdout is the JSON-RPC stream and a stray byte there corrupts the protocol.