VaEmail — email infrastructure for AI agents
MCP server and CLI for VaEmail. Give an agent the ability to send email, authenticate a sending domain, track delivery and diagnose deliverability — without a human reading a dashboard in between.
European infrastructure: servers in Germany, sending through Amazon SES Europe.
Install
# Claude Code
claude mcp add vaemail --env VAEMAIL_API_KEY=swm_your_key -- npx -y vaemail mcp
# Any MCP client
npx -y vaemail mcp
Or as a config block:
{
"mcpServers": {
"vaemail": {
"command": "npx",
"args": ["-y", "vaemail", "mcp"],
"env": { "VAEMAIL_API_KEY": "swm_your_key" }
}
}
}
GET /api/v1/me identifies the tenant, key prefix, scopes and mode without
returning the secret. Contact operations accept a tenant-unique
external_reference for CRM identifiers and upserts.
Use a sandbox API key for integration tests: reads work normally, while every write
is validated in a transaction that is always rolled back. No row, job, email, SMS,
webhook or outbound HTTP call is emitted, and idempotency keys are not stored.
Sandbox write responses carry X-VaEmail-Sandbox: 1 and sandbox: true in JSON
objects. A sandbox response is never evidence that an email was delivered.
Check everything is wired up:
npx vaemail init
Tools
Both vaemail_confirmer_envoi and vaemail_send_email advertise
destructiveHint: true: sending externally is irreversible. The newsletter confirmation
guard applies to immediate sending; drafts and scheduled campaigns keep their existing behavior.
What this is built to prevent
An agent fails differently from a person. It retries, it does not read a dashboard, and it reports success from a 200. The API is shaped around that:
- A send is accepted, not delivered.
vaemail_send_emailreturns 202 and an id. Onlyvaemail_get_messagesays what became of it. The tool descriptions say so, so an agent does not announce a delivery it cannot know about. - Retries do not duplicate. Pass
idempotency_keyand a repeat call replays the first response for 24 hours instead of sending twice. - Errors say what to do. Every failure carries a stable
code, whether it isretryable, and the corrective action with the endpoint that performs it. - Keys are bounded. Scopes, a daily send cap, a cap on recipients per email and an allow-list of sender domains. An agent cannot spend more than the key allows.
- Every call is logged. Which key, which operation, which parameters, which result — message bodies excluded. Autonomous, but accountable.
- Human steps are named as such. Publishing DNS records is one. A record
whose value still contains a placeholder is flagged
publishable: false, because publishing it would break the domain's authentication.
CLI
vaemail init # check config, print the MCP snippet
vaemail capabilities # what the service can do (no key needed)
vaemail send --to [email protected] --subject Hi --html '<p>Hello</p>'
vaemail status 42 # delivery status of a message
vaemail domains:add exemple.fr # declare a domain, print DNS records
vaemail doctor # why is my email not arriving?
vaemail usage # quota and remaining allowance
vaemail mcp # run the MCP server on stdio
Add --json to any command for machine-readable output.
Use it as a library
The same client the MCP server and the CLI run on is exported, so an application can call VaEmail directly. No dependencies.
import { VaEmail } from 'vaemail';
const client = new VaEmail({ apiKey: process.env.VAEMAIL_API_KEY });
await client.send(
{
to: '[email protected]',
subject: 'Your order is on its way',
html: '<p>Tracking number: 1Z999</p>',
},
'order-4711', // idempotency key: safe to replay for 24 hours
);
Errors carry what to do next, not just a status code:
try {
await client.send({ to: '[email protected]' });
} catch (error) {
error.code; // DOMAIN_NOT_VERIFIED
error.retryable; // false
error.pourAgent(); // reason, corrective action, whether to retry
}
Also: capabilities(), health(), validate(), getMessage(),
listMessages(), listDomains(), addDomain(), verifyDomain(),
dnsRecords(), diagnoseDeliverability(), listSuppressions(),
removeSuppression(), usage(), auditLogs().
Connecter Stripe
Utilise une clé Stripe restreinte rk_live_ ou rk_test_, avec Customers en lecture et
Webhook Endpoints en écriture. Une clé secrète sk_ est refusée. VaEmail ne renvoie
aucun secret Stripe : l'état de la connexion expose uniquement les quatre derniers
caractères de la clé et du token du webhook.
Routes REST :
GET /api/v1/integrations/stripe: lire l'état de la connexion ;POST /api/v1/integrations/stripe: créer ou réconcilier le webhook, avec{"api_key":"rk_live_..."}et la portéesettings.write;DELETE /api/v1/integrations/stripe: supprimer le webhook Stripe puis la connexion locale. Action irréversible, à confirmer explicitement par un humain.
Stripe appelle POST /api/v1/webhooks/stripe/{token}. Cette route publique vérifie
Stripe-Signature et ne reçoit pas la clé API VaEmail.
Événements pris en charge et déclencheurs d'automatisation :
A Python SDK with the same surface is available: pip install vaemail
(https://github.com/vaemail/vaemail-python).
Environment
Also available
- OpenAPI specification: https://app.vaemail.fr/openapi.json
- Capabilities endpoint: https://app.vaemail.fr/api/v1/capabilities
- Agent documentation: https://vaemail.fr/agents
Skill and examples
- Agent skill (rules an agent should follow before its first call):
npx skills add vaemail/skills, source https://github.com/vaemail/skills - Examples for curl, Node, Python, Claude Code, Codex, OpenAI Agents, LangChain, CrewAI and n8n: https://github.com/vaemail/examples
- Registry entry:
io.github.vaemail/vaemail(seeserver.json)
Dependencies
None. Node 18+ for the built-in fetch, and nothing else — a package an agent
installs on its own should not pull a dependency tree behind it.
License
MIT