Install Sending in your agent

Sending is not installed: it is a remote MCP server. Nothing to download, no npm package, no process to keep running. Your client needs two things:

Which credential to pick

API key if the client accepts custom headers (Claude Code, Cursor, VS Code, mcp-remote). You create it in Settings › API Keys, pick its scopes there, and see it in plaintext only once. It always points at the same workspace: the right choice for a project or a build machine.

OAuth 2.1 if the client opens a consent screen in the browser (ChatGPT, claude.ai, Claude Desktop). You paste no secret: you authorize with your account and choose the permissions on the consent screen. It is also the only route in clients that cannot send custom headers.

Neither one bypasses the sending guardrails (verified domain, suppression, idempotency, plan quota).

Claude Code

# with an API key
claude mcp add --transport http sending https://sending.dev/api/mcp \
  --header "Authorization: Bearer sk_your_api_key"

Add --scope user to get it in every project, --scope project to write it into the repo's .mcp.json (shared with the team: in that case do not put the key in plaintext). Check with claude mcp list: it must answer ✔ Connected.

By hand, in .mcp.json or ~/.claude.json:

{
  "mcpServers": {
    "sending": {
      "type": "http",
      "url": "https://sending.dev/api/mcp",
      "headers": { "Authorization": "Bearer sk_your_api_key" }
    }
  }
}

Claude Desktop and claude.ai

These go through connectors, so OAuth.

The client registers itself (RFC 7591), so there is no OAuth app to create.

ChatGPT

Custom connectors sit behind developer mode and accept only OAuth or no authentication: there is no way to pass an API key, use consent.

From the API (Responses) the endpoint is passed as a tool, and there an API key works again:

{
  "model": "gpt-5",
  "tools": [{
    "type": "mcp",
    "server_label": "sending",
    "server_url": "https://sending.dev/api/mcp",
    "authorization": "sk_your_api_key",
    "require_approval": "always"
  }],
  "input": "How many emails do I have left this month?"
}

Cursor

In .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "sending": {
      "url": "https://sending.dev/api/mcp",
      "headers": { "Authorization": "Bearer sk_your_api_key" }
    }
  }
}

Then Settings › MCP: the server must show up green with its tools listed.

VS Code (GitHub Copilot)

In .vscode/mcp.json. The key is servers, not mcpServers, and with inputs you type the key on first launch instead of committing it:

{
  "inputs": [
    { "id": "sending-key", "type": "promptString", "description": "Sending API key", "password": true }
  ],
  "servers": {
    "sending": {
      "type": "http",
      "url": "https://sending.dev/api/mcp",
      "headers": { "Authorization": "Bearer ${input:sending-key}" }
    }
  }
}

stdio-only clients (Cline, Windsurf, Gemini CLI)

Clients that do not speak remote HTTP go through mcp-remote, which bridges the two: stdio on one side, our endpoint on the other.

{
  "mcpServers": {
    "sending": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote", "https://sending.dev/api/mcp",
        "--header", "Authorization: Bearer sk_your_api_key"
      ]
    }
  }
}

Without --header the bridge starts the OAuth flow in the browser on first connection.

Any other client

If your agent is not in the list, these parameters are all it takes: URL https://sending.dev/api/mcp, streamable HTTP transport, header Authorization: Bearer <api key or access token>. A client that implements MCP fully finds everything on its own:

• /.well-known/mcp/server-card.json describes server, transport and authentication
• /.well-known/oauth-protected-resource and /.well-known/oauth-authorization-server expose the OAuth metadata (dynamic registration RFC 7591, authorization_code plus PKCE S256)
• /auth.md explains the credentials in prose, for an agent that reads
• /.well-known/agent-skills/ publishes the skills, connect-mcp among them

To hand an agent both REST and MCP in one go, there is a ready-made prompt.

Check that it works

Ask the agent to call get_usage: it is read-only, sends nothing, and answers with the plan, this month's quota and the capacity limits. If you get JSON back, the install is fine.

Permissions

With an API key, the permissions are the scopes you chose at creation, under API Keys.

With an OAuth connection (the client opens the consent screen) you choose the permissions at authorization time: the screen lists the full catalogue, with whatever the client asked for pre-selected. This is needed because MCP clients do not know Sending's scopes and usually ask for identity only: with nothing ticked, the tools answer forbidden: {"need":"..."}.

You can change the permissions of a live connection from MCP Connections, without revoking and reconnecting: they apply from the next call.

If something does not work