What is an MCP server, and how does an agent use one to send messages?

An assistant arranging a delivery notification needs access to the sender, recipient and message content. A connected tool lets it submit those fields without putting your credential into the conversation.

MCP is the protocol connecting the application to those tools.

What does the MCP server provide?

The server gives your AI client tool names, input schemas and callable operations.

Under the MCP tool model, the client discovers tools. It submits structured arguments when the agent selects one.

Bird's tools include sending email, SMS and WhatsApp messages. Setup tools manage sending domains, numbers and sender requirements. Verification tools start a verification. They also check codes. A separate tool advances to another channel.

The available tool list in your client identifies each operation and its required inputs. Check that list when an expected operation is absent.

How do I connect to the hosted server?

Add https://mcp.bird.com as a remote MCP server. Sign in through the browser.

You do not need to install the CLI or supply an API key. Browser sign-in lets the client hold a token with a subset of your permissions. Bird checks each request against the permissions in that token.

Some clients show the server as needing authentication until you start sign-in. A client with only the URL can list tools. Without authentication, a tool call can return 401.

Complete authentication before testing a tool call. The connection guide gives the setup steps for supported clients.

How does an agent send and check a message?

Confirm the recipient has consented to the message and the content meets your applicable policy. Call the channel's send tool. Read the resulting message's delivery events.

For email, the sequence is:

1. Confirm the workspace, sending identity, recipient and content.
2. Call email_send with the required fields from its input schema.
3. Save the returned message ID.
4. Use email_get to inspect delivery status, recipient results and the event timeline.

An accepted send does not prove arrival. A later bounce or rejection needs your agent or application to handle the failure.

A missing sender or permission can prevent the send. Setup and diagnostic tools let the agent inspect those requirements before retrying.

Can I run the server locally?

Yes. Install the Bird CLI. Sign in with bird auth login. Configure your client to launch bird mcp.

The client starts the process. It communicates through the process's input and output streams. It needs bird as the command and mcp as the argument.

The local and hosted servers expose the same toolset. Running the process locally still requires network access for Bird API operations.

Use the local process when you want to manage it yourself. Your client must be able to launch commands. Use the hosted URL when it cannot run a local binary.

What controls which actions the agent can take?

The connected credential limits access. Your AI host controls when tool execution needs confirmation.

Bird annotates destructive tools so the host can recognize them. For example, removing an email suppression is marked destructive because it can resume unwanted sending.

Annotations do not themselves force an approval prompt. Configure your host's policy before allowing changes to production resources.

An agent that can draft and submit an email template can also publish it. A draft step alone does not prevent the agent from calling the submit tool.

WhatsApp template submission sends the template to Meta for review. It does not provide a human approval step for every other action the agent can take.

Voice tools inspect calls. They also issue client credentials. No MCP tool places a voice call. The CLI provides the test-call operation.

For shell-capable agents, the CLI provides JSON output and exit codes. Choose the interface your host can execute. Set its confirmation policy for the intended actions.