AgentSend Skill for OpenClaw

OpenClaw runs skills. AgentSend publishes one. In one install, your OpenClaw agent goes from "I can't send email" to "I have my own @agentsend.io inbox and know how to use it." No API key copy-paste, no signup pre-flight, no config file tinkering.

This page covers how to install the AgentSend skill in OpenClaw, what tools the agent gets, and how the sandbox-to-claim flow works so you only ask a human to intervene at the moment it matters.

Install in One Line

From the OpenClaw CLI, install the skill directly from skills.sh:

openclaw skill install agentsend/agentsend

OpenClaw fetches the skill manifest, installs the @agentsend/mcp runtime, registers the MCP server with your agent, and is ready immediately. On first invocation the skill auto-provisions a sandbox account — no API key required.

What the Agent Gets

After install, the agent has these tools available in every turn:

• create_inbox — Provision a new email inbox with optional display name.
• list_inboxes — List all inboxes. For sandbox accounts, includes quota and claim URL.
• send_email — Send plain text or HTML. Supports CC, threading, and reply headers.
• list_emails — Recent messages in an inbox, filterable by status.
• get_email — Full body and headers of a single message.
• list_threads / get_thread — Thread-grouped conversation view.
• register_webhook — Subscribe to real-time events (requires claimed account).

Sandbox First, Claim When It Matters

The skill's default path is built for instant gratification. On first run the MCP server POSTs a stable device identifier to /auth/auto-provision and gets back an API key tied to a sandbox account. That account can:

• Create 1 inbox
• Send and receive mail freely with other @agentsend.io addresses
• Burn through 50 total sends over up to 7 days

When the agent tries to send to an external recipient, add a second inbox, register a webhook, or exceeds 50 sends, the API returns a CLAIM_REQUIRED error with a claim_url. OpenClaw surfaces the URL to you; you open it, enter a real email, verify a 6-digit code, and the account upgrades to a free-tier claimed account. The API key the agent is already using keeps working — no restart, no re-auth.

{
  "error": {
    "code": "CLAIM_REQUIRED",
    "message": "Sandbox accounts can only send to other @agentsend.io addresses.",
    "status": 402,
    "details": {
      "claim_url": "https://agentsend.io/claim?token=..."
    }
  }
}

Using It: Prompt Examples

Once installed, the agent uses email naturally. A few prompts that work out of the box:

Spin up a support address

"Create an inbox called Support Bot and check if it has any mail."

Fire off a test message to yourself

"Send a test email from my agent inbox to agent-demo@agentsend.io
with subject 'hello' and body 'skill works'. Then check that inbox."

Reply in thread

"Pull the last thread in my inbox and draft a reply explaining
that we've processed the refund."

Why Distribute Via OpenClaw

OpenClaw handles the three things that usually make MCP setup miserable: install, credential storage, and permissions. The AgentSend skill ships with a manifest that OpenClaw reads to know which binary to run, which env vars are required, and which secrets to mask. The agent doesn't see the API key directly — it flows through the runtime.

The net effect: zero friction to "hello, I have an email address" for the agent, zero signup friction for the human, and a well-defined claim moment when the work actually matters.