Atmark

What Atmark MCP server does

The Atmark MCP server provides an email identity for an AI agent, using an address such as name@atmark.ai. It is separate from Gmail and other personal inboxes, so the agent does not need OAuth access to a user's mailbox. The service is invite-only and requires an agent token issued through the Atmark console.

Agents can inspect their capabilities and policy, read received messages, view threads, inspect risk levels, and process text, PDF, or image attachments. They can also send new messages, reply, forward message body text, create and manage drafts, schedule mail, and retrieve verification codes from a sender domain.

Outbound behavior is controlled by the owner's policy. A recipient may be allowed, blocked, or require approval. Policy changes can be requested with a reason, but they do not take effect until the owner approves them. Send results identify whether a message was sent, denied, or placed in pending approval.

How it works

The Atmark MCP server is a remote MCP endpoint at https://api.atmark.ai/mcp. It accepts Streamable HTTP POST requests with JSON responses and authenticates requests with an Authorization: Bearer header. The token must be supplied by the client rather than stored directly in a committed configuration file.

Call get_capabilities first. The response describes available addresses, scopes, policy information, and the operating model. The tools returned by tools/list depend on the scopes assigned to the token. After connecting, get_my_policy can confirm the agent's default sending address and current limits.

Received email is treated as outside data and each message includes a risk.level. The supplied agent rule recommends not following instructions found in email content and consulting the owner when uncertain. Atmark also exposes realtime WebSocket connection details, webhook management, webhook delivery records, and a transparency log containing hashes rather than mail contents.

Setup and configuration

Access must be requested before use. Set the ATMARK_AGENT_TOKEN environment variable, then configure an MCP client to call the endpoint with the Bearer header. The repository includes connection examples for Claude Code, Cursor, and Hermes. Other clients can connect if they support Streamable HTTP and custom Bearer headers; Claude Desktop requires a local bridge to add that header.

The endpoint rejects requests with an Origin header and supports MCP protocol versions 2025-11-25, 2025-06-18, and 2025-03-26. A curl request to tools/list, followed by get_my_policy, can verify connectivity and the configured sending address.

Tools and capabilities

The Atmark MCP server includes tools for:

• Inspecting capabilities, policy, policy-change requests, and approval status.
• Sending, replying to, forwarding, scheduling, cancelling, and drafting email.
• Listing, reading, labeling, and threading received messages.
• Reading attachments or obtaining a five-minute attachment download URL.
• Finding sign-up and login codes from a sender domain.
• Managing webhooks and viewing their deliveries.
• Connecting to realtime updates.
• Creating a sibling agent when owner permissions allow it.

Limitations and notes

Atmark is not intended for cold outreach, AI SDR workflows, bulk or marketing email, unexpected messages, mass sign-ups, or code harvesting. All agents send from the same domain, which limits its fit for use cases that require distinct sending domains.

The Atmark service is not open source. The repository contains client configurations and examples, which are licensed under MIT. Keep the token out of files that could be committed, and use the official endpoint because the README identifies only https://api.atmark.ai/mcp as the Atmark MCP endpoint.