How do I add coss.com to a project with a coding agent?

Give the agent one prompt. It tells the agent to read coss.com/llms.txt, get a development key with npx coss-cli init --agents, and build and test with it. You approve the key in your browser, and nothing the agent sends is delivered until you switch to a production key yourself.

Does coss.com have an MCP server?

Yes. coss-mcp runs locally through npx and gives agents tools to search the docs, find emails and SMS and explain why one wasn't delivered, send, manage suppressions, domains and webhooks, read logs, inboxes and received mail. The docs tools work without an API key.

Does coss.com have an llms.txt?

Yes. coss.com/llms.txt opens with a short brief for agents and lists every docs page, and llms-full.txt has every guide in one file. Every docs page is also Markdown at its URL plus .md.

Is there a Claude Code plugin for coss.com?

Yes. Install it with /plugin install coss --marketplace cosscom/coss. It adds a skill for building with coss.com and the MCP server. The same skill installs into Cursor, Codex and other agents with npx skills add cosscom/coss.

Can an AI agent have its own email address?

Yes. Inboxes give an agent its own address, keep the mail it receives as threads and send its replies from the same address. The MCP server has tools to create inboxes, read threads and reply.

Built for the agent writing your code.

coss.com for coding agents is a one-line setup for Claude Code, Cursor and Codex. Markdown docs, a sandbox key, an MCP server and an Agent Skill.

One prompt

Paste this into Claude Code, Cursor, Codex or any coding agent. It reads the docs, gets a sandbox key and tests what it builds.

Add coss.com to this project: read https://coss.com/llms.txt, get a development key with `npx coss-cli init --agents`, and build and test with it.

From there the agent asks you to approve a development key in your browser, installs the SDK and runs its code against that key. Development keys deliver nothing: emails and SMS get simulated outcomes, events and webhooks. Coding agents has the details.

What the agent gets

Every surface an agent touches is written for it: the docs, the key, the tools and the SDK's results.

Docs as Markdown

/llms.txt starts with a brief (getting a key, how the SDK reports errors, testing webhooks) and lists every page. Every docs page is Markdown at its URL plus .md, and /llms-full.txt has every guide in one file.

A key without copy and paste

npx coss-cli init opens the dashboard, you approve a key for that project (sandbox environments only), and it writes COSS_API_KEY to .env.local or .env and installs coss. Over SSH it prints a link and waits for a code. CLI.

Notes for later sessions

init adds a short coss.com section to the project's AGENTS.md (or CLAUDE.md), so the next session knows where the docs are and how to test. It detects Nodemailer, Better Auth, Auth.js, Laravel, Rails and Django and prints that drop-in.

An MCP server

coss-mcp searches and reads the docs, finds emails and SMS, explains why one wasn't delivered, reads logs, and manages domains, suppressions, webhooks, templates, inboxes and Risk. MCP server.

A plugin and an Agent Skill

The Claude Code plugin adds a skill that loads when a task involves coss.com, and the MCP server. The same skill installs into Cursor, Codex, GitHub Copilot and Windsurf with npx skills add cosscom/coss.

Results agents can read

Every SDK method returns { data, error } and never throws for API or network failures. Errors carry a type, a code, the field's issues and often a hint with a docs URL, so the agent fixes its own request.

The MCP server

coss-mcp is a Model Context Protocol server that runs locally through npx and calls coss.com with your API key.

claude mcp add coss -e COSS_API_KEY=coss_test_… -- npx -y coss-mcp

Ask it "did the welcome email to jane@example.com get delivered?" and diagnose_email answers with a verdict, what happened to each recipient and why, problems with the account and what to do next. get_setup is its first call before writing send code: the key's mode, a from address that will succeed, missing DNS records and one next step.

Docs, with no key: search_docs, read_docs.
Email: diagnose_email, find_emails, get_email, send_email, suppressions, domains and templates.
SMS: diagnose_sms, find_sms, send_sms, numbers and suppressions.
Webhooks and logs: list_webhook_deliveries, retry_webhook_delivery, create_webhook_endpoint, search_logs.
Inboxes and receiving: create_inbox, get_inbox_thread, reply_to_inbox_message, list_received_emails.
Hooks, Forms and Risk: send_hook_message, list_form_submissions, assess_risk.

Email content, received mail, bounce messages and webhook responses are written by other people: the server labels them untrusted and tells the agent to treat them as data, never as instructions. Without COSS_API_KEY it reads only a development key from the project's env file, never a coss_live_ one.

Safe to retry, safe to test

Agents retry and run code repeatedly. coss.com makes both harmless.

const { ,  } = await ..(
  {
    : "Acme <hello@acme.com>",
    : .,
    : "Welcome to Acme",
    : "Thanks for signing up.",
  },
  { : `welcome-${.}` }, // a rerun sends nothing twice
);

if () .(., .); // what to change, not a stack trace

The SDK sends an Idempotency-Key with every POST and reuses it when it retries, so a retried send goes out once. coss listen forwards sandbox webhooks to localhost, and coss trigger sends a sample event of any type, so an agent can test its webhook handler end to end.

Email for the agents you build

Coding agents build with coss.com; the agents you ship can use it too.

Inboxes give an AI agent its own email address: mail arrives as threads with a webhook for each message, and the agent replies from the same address. Risk can classify what a message says before your agent acts on it.

Questions

How do I add coss.com to a project with a coding agent?: Give the agent one prompt. It tells the agent to read coss.com/llms.txt, get a development key with npx coss-cli init --agents, and build and test with it. You approve the key in your browser, and nothing the agent sends is delivered until you switch to a production key yourself.
Does coss.com have an MCP server?: Yes. coss-mcp runs locally through npx and gives agents tools to search the docs, find emails and SMS and explain why one wasn't delivered, send, manage suppressions, domains and webhooks, read logs, inboxes and received mail. The docs tools work without an API key.
Does coss.com have an llms.txt?: Yes. coss.com/llms.txt opens with a short brief for agents and lists every docs page, and llms-full.txt has every guide in one file. Every docs page is also Markdown at its URL plus .md.
Is there a Claude Code plugin for coss.com?: Yes. Install it with /plugin install coss --marketplace cosscom/coss. It adds a skill for building with coss.com and the MCP server. The same skill installs into Cursor, Codex and other agents with npx skills add cosscom/coss.
Can an agent test its code without sending real email?: Yes. The development key coss-cli init writes works with every API and delivers nothing. Emails and SMS get simulated outcomes, events and webhooks, and coss listen forwards those webhooks to the agent's local handler.
Can an AI agent have its own email address?: Yes. Inboxes give an agent its own address, keep the mail it receives as threads and send its replies from the same address. The MCP server has tools to create inboxes, read threads and reply.

On the same key, SDK, webhooks and logs: nothing more to set up.

Start with Coding agents today.

Development keys record every request and event and deliver nothing, so you can build and test before you verify a domain. Hand your coding agent this prompt:

Add coss.com to this project: read https://coss.com/llms.txt, get a development key with `npx coss-cli init --agents`, and build and test with it.