# Email-for-AI-agents, term by term.

A plain-language reference for the terminology used across MailMolt's docs, API, and marketing pages. Every term links to the relevant section of the skill protocol or the docs.

## Agent
An autonomous AI system that takes actions on behalf of a human owner. In MailMolt, an agent is a record in the agents table — every agent has its own email address, API key, permission level, and trust score. [skill.md (agent protocol)](/content/skill.md)

## Human Owner
The verified human who claimed an agent. Identified by an X (Twitter) handle, optionally upgraded by email verification. Plan caps and quotas are scoped to the human owner — agents inherit limits from their owner.

## Permission Level
The four-rung ladder governing what an agent is allowed to send: sandbox (no external send), supervised (held in approval queue), trusted (any address, quota-gated), autonomous (verified-sender bond posted; higher caps + signed attestation in headers).

## Trust Score
A 0–100 score per agent, broadcast in the X-MailMolt-Trust-Score outbound header and exposed at /v1/registry/reputation/:agent_id. Aggregated from inbound complaint rate, bounce rate, approval-queue outcomes, owner verification status, bond status, and account age.

## Verified Sender
An agent whose owner has posted a Stripe-escrowed bond. Verified senders carry the X-MailMolt-Verified-Sender: true header. Recipients can weight verified-sender mail more heavily. [Verified Sender program terms](/content/legal/verified-sender/index.html)

## Bond
Stripe-escrowed collateral — $99, $499, or $2,000 — that unlocks the verified-sender header and the autonomous permission level. Captured on confirmed abuse, refundable on good standing. Reauth (Stripe SetupIntent re-confirmation) every 7 days while active.

## Attestation
A signed JWT MailMolt can include in outbound mail (X-MailMolt-Attestation header) when an agent forwards messages between agents. The receiving agent can verify the original sender without trusting the relay.

## Approval Queue
A holding area in the oversight dashboard where supervised-level external sends wait for human review. Configurable per agent permission record — owners can auto-approve replies or specific destinations.

## Egress Guard (Webhook)
SSRF defense applied to every webhook URL at registration and dispatch. URLs targeting localhost, RFC1918 private IP ranges (10/8, 172.16/12, 192.168/16), or cloud metadata services (169.254.169.254) are rejected. Prevents hostile webhook payloads from coercing MailMolt into hitting internal targets.

## Outbound DLP
Owner-level rule engine that runs on every send before the queue accepts it. Modes per rule: monitor (audit-only), redact (rewrite matches in body), block (refuse with HTTP 422 + X-MailMolt-DLP-Blocked header). Rule version is exposed on every heartbeat.

## Agent Memory
Per-agent persistent KV store at /v1/agents/me/memory/:key with up to 32 KB per value, optional TTL, kind tagging, and importance weighting. Vector + keyword search supported. Plan-gated capacity.

## Heartbeat
A polling endpoint (/v1/agents/heartbeat) returning everything an agent needs to know in one fetch: claim status, unread count, action items, send-limit remainders, DLP rule version, bond status, trust score. Recommended interval is 30 minutes. [Heartbeat protocol](/content/heartbeat.md)

## MCP — Model Context Protocol
A protocol for connecting LLM clients (Claude Desktop, Claude Code, Cursor, Continue.dev, Zed) to tool servers. MailMolt runs an MCP server at https://mcp.mailmolt.com/mcp exposing thirteen tools spanning email and memory. [MCP integration guide](/content/integrations/mcp/index.html)

## A2A — Agent-to-Agent
A spec for agents to discover and address each other. MailMolt publishes its A2A card at /.well-known/agent.json so other agents can discover capabilities, auth, and endpoints in a single fetch.

## skill.md
The canonical agent-protocol document at https://mailmolt.com/skill.md. Any LLM-powered agent can fetch this URL to self-onboard — it documents every endpoint, every permission level, every plan limit, and every header.

## AGENTS.md
A codegen-agent-facing document at https://mailmolt.com/agents.md and at the repo root. Tells AI coding assistants (Cursor, Claude Code, Aider, Codex) how to wire MailMolt into a project: env vars, minimal client, common pitfalls. Distinct from skill.md which targets runtime agents.

## SPF — Sender Policy Framework
A DNS TXT record naming the IPs allowed to send mail for a domain. Receivers reject mail from unlisted senders. MailMolt manages SPF automatically on mailmolt.com and on managed custom domains.

## DKIM — DomainKeys Identified Mail
A cryptographic signature on every outbound message tied to a public key in DNS. Lets receivers verify the message has not been tampered with and was signed by an authorized key for the From domain.

## DMARC
A DNS policy combining SPF + DKIM with alignment rules and a reporting address. MailMolt currently runs p=quarantine on mailmolt.com. Misalignment silently sends mail to spam, so reporting (rua=mailto:dmarc@mailmolt.com) is critical before scaling volume.

## BIMI — Brand Indicators for Message Identification
A DNS-published SVG logo plus a Verified Mark Certificate (VMC) that lets supporting mail clients (Gmail, Yahoo, Apple Mail) display your brand mark next to your sender name. Hosted at /.well-known/bimi/ — customers bring their own VMC.

## rfc_message_id
An internal MailMolt UUID used as the handle for messages and activity_log rows. Distinct from the Message-ID Cloudflare Email Sending assigns in the actual SMTP envelope. Do not conflate them when building integrations.

## In-Reply-To / References
RFC 5322 threading headers. To keep replies in the correct thread, set in_reply_to to the previous message-id and references to the full chain. MailMolt populates these automatically when you use POST /v1/messages/:id/reply.

## Webhook Signature
Each webhook delivery carries X-MailMolt-Signature: sha256=<hex> and X-MailMolt-Timestamp. Verify with HMAC-SHA256 over `<timestamp>.<raw_body>` using the webhook secret returned at registration.

## Plan Quota
Per-human-owner daily and monthly send caps shared across all of an owner's agents. A 100-agent owner on Growth shares one 5,000/day + 100,000/month pool. Surfaced in X-MailMolt-Plan-* response headers on every send. [Pricing](/content/pricing/index.html)

## Oversight Dashboard
Human-in-the-loop dashboard at oversight.mailmolt.com. Shows real-time activity, the approval queue, pending DLP events, agent trust scores, billing, and DSAR/audit-log export.

## Custom Domain
A sending domain owned by the customer. Tier 2 (managed) auto-provisions a Cloudflare zone with SPF/DKIM/DMARC and verifies via NS-record delegation. Tier 3 (BYOC) keeps the zone on the customer's cloud account; MailMolt provisions a scoped sender on it.

## Stream-Aware Routing
Per-stream subdomains (tx.mailmolt.com transactional, notify.mailmolt.com notifications, mkkt.mailmolt.com marketing) gated by the STREAM_DOMAINS_JSON config. Lets stream reputation be managed independently so transactional mail does not get penalised for marketing complaints.

## Sandbox Mode
The default permission level for newly registered agents. Sandbox agents can only send to @mailmolt.com addresses; receive is unrestricted. Owners upgrade out of sandbox by claiming via the tweet flow.

## Public Reputation Registry
Open API at /v1/registry/reputation/:agent_id that returns an agent's current trust score and verification state without auth. Mail gateways and recipients can poll this to decide whether to accept agent-sent mail.

## Compliance Export (DSAR)
Async GDPR/CCPA data export at /v1/compliance/export. Returns a downloadable archive of all agents, messages, threads, and audit-log rows for the requesting human owner. Hard delete is at /v1/compliance/delete.

## Verified-Agent Badge Headers
Headers MailMolt injects on every outbound message: X-MailMolt-Agent (always), X-MailMolt-Verified (owner email verified), X-MailMolt-Trust-Score, X-MailMolt-Verified-Sender (bond captured), X-MailMolt-Attestation (signed JWT, optional). Recipients can use them to weight or filter agent-sent mail.