Agent Skills for email
A skill is a markdown file that teaches an agent one job end to end: the calls, their order, and what to do when one is refused. These five cover the work an agent actually has to do with email, and they are the same files our own MCP server hands out.
Served at /.well-known/agent-skills/ · discovery spec 0.2.0 · free to copy
The five skills
Connect an AI agent to Sending over MCP
Connect to Sending's MCP server and get the right permissions. Use this before anything else, and whenever a tool answers forbidden while the connection looks fine.
Let an AI agent send its first email
Verify a sending domain and send your first transactional email with Sending. Use this when you need to send a single email, or when a send is refused because the domain is not verified.
Have an AI agent run an email campaign
Build and send an email campaign to a list or a segment with Sending. Use this for bulk sends, newsletters and announcements, not for single emails.
Have an AI agent build a marketing automation
Create, fix and activate a graph automation (welcome, abandoned cart, follow-up) with Sending, enroll contacts by hand, and read the logs to find out why a contact never got a message. Use this for recurring sends triggered by an event, and before changing automations that are already live.
Give an AI agent its own email inbox
Give an AI agent a real email inbox with Sending: create the inbox, read the threads it receives, and reply. Use this when the agent has to receive mail, not just send it.
How an agent finds them
The index is generated from the files themselves, and the digest is the sha256 of the exact bytes served. That is the part worth checking in someone else's implementation too: an index maintained by hand starts describing skills that no longer say what it claims, and nothing breaks when it does.
curl https://sending.dev/.well-known/agent-skills/index.json
{
"version": "0.2.0",
"skills": [
{
"name": "connect-mcp",
"description": "Connect to Sending's MCP server and get the right permissions...",
"url": "/.well-known/agent-skills/connect-mcp/SKILL.md",
"digest": "sha256:..."
}
]
}Each file is also readable as a page here, and every page answers in markdown on Accept: text/markdown or with a .md suffix.
What a skill will never do
Teach an agent to get around a guardrail. Sending refuses a send from an unverified domain, drops addresses on the suppression list, requires an idempotency key, applies per-key volume and spend caps, and writes every tool call to an audit log. All of it is enforced on our servers. A skill that promised otherwise would just be describing a product that does not exist.
Questions
What is an agent skill?
A skill is a markdown file that teaches an AI agent how to do one job with a product: what to call, in which order, and what to do when it fails. It is not documentation for humans that an agent happens to read. It is written for the agent, and it is loaded into the model's context when the job comes up.
How does an agent find them?
They are published at /.well-known/agent-skills/index.json, following the skills discovery spec. The index lists each skill with its name, description, URL and a sha256 digest of the exact bytes served, so a client can cache one and know when it has changed.
Do I have to install anything?
No. The skills are plain files served over HTTP, and the tools they describe live on our MCP server at https://sending.dev/api/mcp. There is no package to install and no runtime of ours on your machine.
Can I copy one into my own agent?
Yes. Copy the markdown, keep it, edit it. They are written to be pasted into a system prompt or dropped into a skills folder. The digest in the index tells you when ours has moved on.
What does a skill assume about permissions?
That the agent has an API key or an OAuth grant with the right scopes, and that anything destructive stays gated server-side. A skill never tells an agent how to get around a verified-domain check, a suppression list or a spend cap, because it cannot: those are enforced on our servers, not in the prompt.
